<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Ankit Singla on Medium]]></title>
        <description><![CDATA[Stories by Ankit Singla on Medium]]></description>
        <link>https://medium.com/@stub_AS?source=rss-88a9a115d027------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*trDaOzSobBz81uz92rCi0A.jpeg</url>
            <title>Stories by Ankit Singla on Medium</title>
            <link>https://medium.com/@stub_AS?source=rss-88a9a115d027------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 02:41:12 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@stub_AS/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Where’s the not-fluffy case for crypto?!]]></title>
            <link>https://medium.com/@stub_AS/wheres-the-not-fluffy-case-for-crypto-e3ff71c6c2a0?source=rss-88a9a115d027------2</link>
            <guid isPermaLink="false">https://medium.com/p/e3ff71c6c2a0</guid>
            <category><![CDATA[crypto]]></category>
            <category><![CDATA[web3]]></category>
            <dc:creator><![CDATA[Ankit Singla]]></dc:creator>
            <pubDate>Wed, 02 Feb 2022 15:09:47 GMT</pubDate>
            <atom:updated>2022-02-02T15:09:47.828Z</atom:updated>
            <content:encoded><![CDATA[<p>If your online experience is still untouched by “shilling” for crypto and sniping against it, I guess you don’t need a <a href="https://www.cnbc.com/2021/08/23/people-are-paying-millions-of-dollars-for-digital-pictures-of-rocks.html">virtual rock</a> to hide under. Despite finding many of the technical ideas interesting, I’ve not been able to come up with many concrete use cases for crypto beyond speculation and bragging rights.</p><h4>Cryptocurrencies: unreliable at best, energy-waste at worst</h4><p>Cryptocurrencies are <a href="https://www.bloomberg.com/graphics/2021-bitcoin-volitility/">highly volatile</a>, <a href="https://www.vox.com/recode/2021/5/18/22441831/elon-musk-bitcoin-dogecoin-crypto-prices-tesla">easily manipulated</a>, <a href="https://www.wsj.com/articles/crypto-and-its-many-fees-what-to-know-about-the-hidden-costs-of-digital-currency-11639825202">have high transaction costs</a> for small transactions, often have <a href="https://www.statista.com/statistics/793539/bitcoin-transaction-confirmation-time/">high transaction latency</a>, and are not widely accepted as tender. Worse, BitCoin and other proof-of-work cryptocurrencies require huge amounts of electricity that <a href="https://www.wired.com/story/kazakhstan-cryptocurrency-mining-unrest-energy/">wreak havoc with power grids at country scale</a>. Proof-of-stake coins address some, but not all of these maladies. A variety of alt-coins exist, many of which are (at least) not technologically inferior to BitCoin, underscoring that the perceived value of these crypto coins is <em>not</em> tech-driven, but derives largely from its uptake. This is true of fiat currencies as well, but this is not a winning counterpoint given that within a country, its fiat currency has full uptake, and at least for stable governments, trust in their backing of their currency is global. Even if one were to buy that crypto is great in unstable countries with crazy inflation, is the rest of the globe’s speculation on it as an asset useful?</p><h4>Crypto is replete with scams</h4><p>There’s a well-maintained <a href="https://web3isgoinggreat.com/">running list</a>, so I don’t have to bother. It’s worth noting though that these failings aren’t limited to the fringes of crypto. Instead, some of the most outrageous ones have resulted in key companies in the crypto space, like Tether, being forced into <a href="https://www.cftc.gov/PressRoom/PressReleases/8450-21">huge settlements</a> with regulators. Further, the dependence of crypto valuations on sheer belief, as noted above, often colors the whole space PonziRed.</p><h4>Crypto is unlikely to achieve many of its goals</h4><p>There’s no consensus (ironic?) on what the goals of crypto precisely are. But it seems very unlikely to meet some of the most hyped ones. Take “decentralization” for instance. It’s not entirely clear what type of decentralization advocates are talking of, but if it’s about control by a few large corporations and platforms, well, that ship seems to be sailing away, with massive consolidation in crypto exchanges and NFT platforms. An <a href="https://moxie.org/2022/01/07/web3-first-impressions.html">excellent overview</a> of several aspects of the current crypto/Web-3 ecosystem delves into this and other issues in great detail.</p><p>On “digital ownership”, crypto isn’t headed towards one global blockchain for all use cases, so for any digital assets, ownership claims made on different chains have to be reconciled somehow, likely by centralized third parties like NFT markets, and based on traditional tools and mechanisms for enforcing (intellectual) property rights. These centralized entities may also breach the “censorship resistance” argument for crypto.</p><p>With respect to “opportunity for everyone”, the same <a href="https://a16z.com/portfolio/">venture</a> <a href="https://app.dealroom.co/investors/accel_partners/portfolio">capital</a> <a href="https://app.dealroom.co/investors/tiger_global">firms</a> that funded the largest “Web 2” companies are some of the key stakeholders in crypto, and crypto wealth is as <a href="https://www.frontiersin.org/articles/10.3389/fbloc.2021.730122/full">unequally distributed</a> as fiat, or worse.</p><p>For “shared governance”, other than things that can be resolved entirely on-chain (e.g., if X happens, send Y money to Z address), enforcement relies on trusting the originators/executors of the project. A variety of cooperatives work this way already, with greater regulatory protection.</p><p>The libertarian goal of a reduced government role in finance is also facing regulatory headwinds, for good reason, as I discuss below.</p><h4>“But the technology and ecosystem will keep improving …”</h4><p>Sure, but the strongest criticisms of crypto are ones that even a new breakthrough in the technology for distributed consensus won’t solve. These issues stem from user expectations and necessary regulation.</p><p>First, the reversibility of transactions is essential to correcting errors and fraud protection — this is what our fiat ecosystem relies on to resolve disputes. (Along similar lines, for non-currency projects, data deletion may be essential in certain cirumstances, e.g., removal of hate speech, child-sex abuse materials, and for compliance with laws like Europe’s GDPR.) Second, the tracking of large financial transactions helps track and prevent money laundering, tax evasion, and other crime. Third, modern banking ultimately fulfills many consumer needs, and similar needs arise with crypto:</p><ol><li>Even most tech-savvy folks wouldn’t want to manage crypto-keys to secure their life’s savings. So ultimately, it’s likely that cryptocurrencies will need institutions, much like banks, that provide custodial services.</li><li>There are several high-profile cases of bugs in crypto smart contracts already. Businesses will need insurance from losses of this nature.</li><li>Who will manage reporting for regulatory compliance such as taxation, and for logging large transactions for regulatory review?</li></ol><p>All of this points to the creation of centralized entities that mirror the fiat ecosystem. Ultimately, the frictions and controls in the fiat economy that crypto communities detest are likely to reappear with (needed) regulation.</p><h4>Crypto is attracting a huge amount of talent</h4><p>The primary reason I stop short of espousing the “it’s all a scam” takes is that this seems like an arrogant position given the huge investments of not just money, but far more importantly, of well-respected talent, including folks I personally know to be well-intentioned, solid researchers and engineers. Several crypto projects like Ava, Algorand, ZCash, Chainlink, and Oasis Labs are rooted in academia, with acclaimed researchers playing key roles in them. Others have huge research teams behind them, with a number of PhDs in distributed systems, cryptography, and security; consider the rosters of the teams at <a href="https://dfinity.org/foundation#team">Dfinity</a> and <a href="https://research.protocol.ai/people/">FileCoin</a> for instance.</p><p>The few non-fluffy (i.e., not dependent on ill-defined buzzwords like “decentralized”, and “distributed ownership”) use cases presented so far seem grossly insufficient in relation to the investment in crypto. Are massive, instant, unregulated, cross-border money transfers even a net positive for the world, given their potential for abuse in ransomware and money laundering? There are already good arguments against <a href="https://www.bloomberg.com/news/newsletters/2021-11-12/crypto-in-video-games-is-all-the-rage-but-why">gaming</a> and <a href="https://www.youtube.com/watch?v=YQ_xWvX1n9g&amp;t=4851s">digital art markets</a> being killer use cases, and the reception by gamers (not just gaming companies) has been <a href="https://www.slashgear.com/ubisoft-exec-inexplicably-blames-gamers-for-dire-nft-reception-28709496/">quite hostile</a>, and by artists, <a href="https://www.polygon.com/22327806/nft-artists-online-theft-non-fungible-token">at best mixed</a>. Various proposals that seek to put real-world data on a blockchain all seem to suffer from the same issue: verifying that the interfaces with the real world are not corrupted, i.e., the data is entered truthfully. The only exceptions are fully on-chain systems, which, unfortunately, seem limited to cryptocurrency.</p><p>Why are so many capable folks investing their time into something that doesn’t seem to solve real problems? I hope that one or more of them will make a <strong>concrete</strong>, affirmative case for crypto, i.e., outline at least a handful of use cases that: (a) have non-speculatory value, beyond bragging rights; and (b) couldn’t be better served by non-crypto technologies. Please don’t say “flash loans” or “smart contracts” without telling us the end goals — for use in what? Until someone makes a clear case, I remain very skeptical that this amounts to anything more than a few niche applications. I would be happy to be persuaded otherwise — every nerd wants to play with cool new tech!</p><p>Another question on my mind: if crypto fails, or succeeds only in niches, what do the numerous well-heeled startups pivot to? Microtransactions for various e-businesses? Internet-native cooperatives without blockchain? What else?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e3ff71c6c2a0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Logging research ideas]]></title>
            <link>https://medium.com/@stub_AS/logging-research-ideas-26192626632c?source=rss-88a9a115d027------2</link>
            <guid isPermaLink="false">https://medium.com/p/26192626632c</guid>
            <category><![CDATA[research-idea]]></category>
            <category><![CDATA[ideas]]></category>
            <category><![CDATA[computer-networking]]></category>
            <category><![CDATA[research]]></category>
            <dc:creator><![CDATA[Ankit Singla]]></dc:creator>
            <pubDate>Fri, 21 Jan 2022 21:18:54 GMT</pubDate>
            <atom:updated>2022-01-22T09:39:40.975Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This is part of a series of write-ups I’m working on, drawing on my notes and materials over the last ~7 years. I’m </em><a href="https://twitter.com/stub_AS/status/1469283183132876809"><em>leaving academia</em></a><em> for an exciting industry opportunity after not getting tenure. The research ideas I’m putting out are open for anyone to build on.</em></p><figure><img alt="Light bulbs hanging from a ceiling, with some of them lit." src="https://cdn-images-1.medium.com/max/1024/1*mV2qb6ThKJXpSiE17GK-YA.jpeg" /><figcaption>Quite a few duds in there, and there’s no shame in that :-)</figcaption></figure><p>I am posting a collection with my research ideas in the form of a “<a href="https://drive.google.com/file/d/1l2l2gM6sRhMI1F7UynNBIpCq-f0nal0b/view?usp=sharing">log</a>” I maintained over roughly 5 years (2016-2020). This public version excludes three types of ideas:</p><ul><li>Ideas which I discussed with collaborators and are being pursued, because these I now view as joint work, and am not free to share alone.</li><li>Ideas that already led to papers, because it’s in bad taste to lay claim to ideas for work that ultimately took place with other collaborators.</li><li>Ideas that were “on my main path”, because I kept separate notes for my main directions, e.g., you’ll see very little here on my <a href="https://medium.com/@stub_AS/research-problems-in-satellite-based-broadband-%C3%A0-la-spacex-starlink-6f5dfe6d2756">satellite direction</a>.</li></ul><p>Thus, this log primarily contains ideas I thought could lead to new directions for my group that didn’t get explored yet, and things I wanted to learn more about. As you might expect, ideas of every flavor are in there:</p><ul><li>Sensible ideas I liked but didn’t pursue as they would probably have low impact. See, for example, #37.</li><li>Sensible ideas that my group wasn’t well-positioned to pursue, and others would have scooped us to good execution. See, for example, #30.</li><li>Nonsensical ideas that were poorly conceived, but seemed fine to me at the time. I highly recommend keeping crap ideas around too, because sense and nonsense are sometimes separated from each other only by time. See #9, #15, #25, and #66 — there are many, but need I suffer more?</li><li>Ideas which haven’t cleanly sorted in my mind into sensible/nonsensical. Obviously, based on the previous point, <em>time</em> is at fault! See #13, #18.</li><li>Half-baked thoughts and questions masquerading as “ideas”, e.g., #20.</li><li>Excellent ideas … which had already been pursued by others. See #19.</li><li>Concrete, potentially high-impact ideas I just didn’t get to — these are ones I would have probably pursued, had I continued on my trajectory. See, for instance, #109, #110, #114, and #116. (Not sure whether that’s recency bias, or were my ideas improving over time?)</li></ul><h4>Every researcher should have an idea log</h4><p>I highly recommend the practice of maintaining an idea log. I found idea logging <strong>extremely</strong> motivating and useful for several reasons:</p><ul><li>It gave me an outlet for every stray idea I had, and sometimes, poorly framed or half-baked things met other half-baked things in this log and made each other whole. (Is this from some embarrassing movie?)</li><li>Every time I felt like I hadn’t put down new things in the log for awhile, I felt somewhat pressured to do so, and this prompted more reading, more thinking, and a greater urgency in finding new ground to explore.</li><li>When I lacked for inspiration, scrolling back on my log gave seeds for new thinking, and also reminded me that the bar for what I considered an interesting idea was quite low :)</li><li>I consciously thought at talks and after conversations with colleagues what I could think about and take back to my idea log.</li><li>Putting things down in the idea log often forced me to do some basic searching around to try and add the closest related thing I could find. This also told me whose work I should be following some similar threads, and gave me good starting points for conversations with a wide variety of folks.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=26192626632c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Publishing and sharing authorship]]></title>
            <link>https://medium.com/@stub_AS/publishing-and-sharing-authorship-646ffdcb797?source=rss-88a9a115d027------2</link>
            <guid isPermaLink="false">https://medium.com/p/646ffdcb797</guid>
            <dc:creator><![CDATA[Ankit Singla]]></dc:creator>
            <pubDate>Thu, 20 Jan 2022 11:56:32 GMT</pubDate>
            <atom:updated>2022-01-20T11:56:32.569Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This is part of a series of write-ups I’m working on, drawing on my notes and materials over the last ~7 years. I’m </em><a href="https://twitter.com/stub_AS/status/1469283183132876809"><em>leaving academia</em></a><em> for an exciting industry opportunity after not getting tenure.</em></p><p>I wrote a version of this a while back for my then PhD students, but perhaps it helps others too. This is in the context of the most competitive <strong>Computer Systems</strong> conferences, with 5+ reviews for each accepted paper, and ~15% acceptance rate. The value of papers and authorship, and what counts as a good record, varies across disciplines and across institutions, and above all, your own circumstances, goals, and expectations of yourself.</p><figure><img alt="“The Massacre of the Innocents” by Rubens (1638) — a renaissance painting showing a biblical massacre." src="https://cdn-images-1.medium.com/max/1024/1*Opy9cjFMJpKKiu7btHcFEg.jpeg" /><figcaption>A program committee meeting in progress: most of your papers are doomed (this time). Collaborative work has the additional perk of shared commisseration.</figcaption></figure><h4>What counts as publishing success over your PhD?</h4><ul><li>In strictly publication terms, your goal should be building a portfolio with a mix: some first-author and (more?) non-first-author work. If possible, try to keep the topics you contribute to somewhat coherent, so that your contribution is identifiable even if you’re not the first author. (Higher-level goals are better, e.g., having <strong>one</strong> super well-recognized project is great!) There’s something to say for doing multiple flavors of work, so you learn more, but try to balance these objectives (learning more vs. using what you know) — it’s a standard explore vs. exploit problem.</li><li>Different portfolios have different strengths, and are read accordingly by others in the field: e.g., finishing with 3 top first-author papers and nothing else is good, but 2 first-author papers plus 2–3 non-first-author ones also make a good impression. There’s not necessarily a good ordering between those two portfolios, assuming these papers are of similar quality.</li><li>By the end of your PhD journey, as long as you have some first-author work, and a reasonable portfolio overall, you’ll fall into the “productive” bucket. Not checking either box, e.g., no first author work, or only 1 or 2 pieces of work (unless these end up being well-known) may not even suffice for graduation. Again, these expectations are in line with tier-1 research-institutions doing research on computer systems — please ask your mentors for recalibration per your circumstances. (I’ve noticed how systems students get disheartened seeing their ML colleagues pile up papers.) Note that there are other means of impact, e.g., building software/systems that are widely used. Also, note that paper acceptances involve way too much randomness* to tie your self-worth to them.</li><li>Author ordering does matter, so you should advocate for yourself, while being fair to others. For papers with only 2–3 PhD students, it matters less than one might think. Contrary to some students’ perception, it’s not that people perceive the first author to have done 90% of the work and other authors as nothing — especially in computer systems, there is a recognition that building and evaluating something often requires significant effort from multiple people. What you and the senior authors in such collaborations communicate about your role will matter.</li><li>Going beyond “productive” to “highly productive” requires either expansion of first-author efforts, or more collaboration with a bigger set of non-first-author work. I <em>think</em> most experienced researchers will agree that the second option is easier. (This was also my path — as a grad student, I wrote more papers as second or third author than as first author.)</li><li>Another big point in favor of collaborative work, even if it sometimes feels more important to do your own thing: it can be extremely enjoyable. Instead of slogging alone, you have smart coworkers to bounce ideas of, come up with something new, detect problems early, etc.</li><li>An under-rated aspect is the social one — both on your own and through your advisor. People will ask your advisor about your performance informally (“how are your students progressing? which student is leading your work on X?”) and formally (letters). If you are in the “productive” or better category, this aspect could have a large impact: how do you, your advisor, and other letter writers present your accomplishments?</li></ul><h4>What actionable advice could you draw from the above?</h4><ul><li>Try and clearly “own” some of your research portfolio.</li><li>Do not worry <strong>too much</strong> about authorship in shared work. Rather, try to collaborate as much as possible. If you’re a bit further into your career, look back and see which papers wouldn’t exist at all without collaboration.</li><li>If you have ideas you don’t have time for, let others lead them, and at least get them a home. It’s better to move science forward and get <em>some</em> credit.</li><li>If possible, before the first line of code is written, or the first sentence of a draft, be clear on who the owner of the work is. Then discuss expectations for effort. Note that if you decide on being second or third author, it doesn’t mean you can phone it in — authorship itself should always entail substantial work. Be open to giving others more credit, e.g., roping in another author as shared-first, if they greatly exceed the set expectations. But as to-be second or third author yourself, don’t expect this, and plan your time and priorities accordingly.</li><li>When differences of opinion arise, be generous and give others the benefit of doubt. It’s really important to communicate with your advisor in clear terms what your expectations are, and when you feel that the credit apportioned to you is unfair. The best time to talk about this is well before a paper is finalized, e.g., at the time of submission.</li><li>Advocate for yourself, within reason. Weigh the costs (are the personalities involved such that there’s high potential for acrimony? could it hurt you in getting good letters? hurt future collaboration?) against the benefits. Occasionally, even if you know you definitely deserve more, the costs may be too high to pursue this. You can and should approach the topic, but if this doesn’t move things in your favored direction, consider carefully if you want to push more. (Maybe get advice from more than one person!) Also note that differences of opinion are real; most people are not trying to rob you, their accounting of work just happens to be different or/and wrong. But you will also occasionally be the one with the wrong perception, so allow for a healthy amount of self-doubt. As for my own experience, a couple of times in the past, I thought I deserved more, but I didn’t pursue it for various reasons. Despite being fairly lax in my approach to authorship, so far, I got completely screwed-over only once; and even in that case, I must have gotten a stellar letter. So … not <em>too</em> bad.</li></ul><h4>* Randomness in paper acceptances</h4><p>At this point, it’s well known that there’s <a href="http://inverseprobability.com/talks/notes/the-neurips-experiment.html">somewhat high randomness</a> in the reviewing process at top conferences in Computer Science. What might this mean for a particular graduate student?</p><p>Say you submit 4 times to venues with 15% acceptance rates. Let’s further suppose that your work, being much cooler and of higher quality than the typical submission, has a higher chance of acceptance at 33%. (I have worked with some of the best-known figures in Computer Networking, and a 1-in-3 accept rate for one’s paper submissions would be considered quite good.) The chance that you get <em>zero</em> acceptances is a disturbingly high 1-in-5.</p><p>Good mentorship can and should help you both hedge your bets in terms of flavor and positioning of your work, and to improve your papers across successive submissions, to avoid such worst-case outcomes. Nevertheless, we should expect that some careers will be derailed by randomness to varying degrees. The zero acceptances case is particularly stark, but whether one arrives at the faculty job market with 5 top publications versus 3 can come down to a few unreasonable or grumpy reviewers, and probably happens way more often than it should.</p><p>The role of “luck” is something most faculty accept. It’s thus good for everyone involved to show compassion by remembering that while most successful people work hard, not everyone who works hard succeeds. This is <em>especially</em> true when the measure of success is as weak as paper acceptances. Be persistent in your efforts, but also try and work around the limits of papers by socializing your work through talks, or presenting at posters and workshops.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=646ffdcb797" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Network design under uncertainty]]></title>
            <link>https://medium.com/@stub_AS/network-design-under-uncertainty-891470040d07?source=rss-88a9a115d027------2</link>
            <guid isPermaLink="false">https://medium.com/p/891470040d07</guid>
            <dc:creator><![CDATA[Ankit Singla]]></dc:creator>
            <pubDate>Mon, 10 Jan 2022 12:48:19 GMT</pubDate>
            <atom:updated>2022-01-10T12:48:19.527Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This is part of a series of write-ups I’m working on, drawing on my notes and materials over the last ~7 years. I’m </em><a href="https://twitter.com/stub_AS/status/1469283183132876809"><em>leaving academia</em></a><em> for an exciting industry opportunity after not getting tenure. The research ideas I’m putting out are open for anyone to build on. The caveat is that my background work on many of these, including a full review of the relevant literature, is incomplete — this is just a snapshot of where I happened to be in my thinking along these lines.</em></p><p>In designing and expanding networked systems, operators are often faced with uncertainty in what workloads their system might serve — they may only have an imprecise forecast for future traffic. This imprecision manifests as an “uncertainty tax” on network design by requiring that the operator overprovision their network substantially to cover for possible bad cases. However, in some contexts, there are interesting research opportunities for reducing this uncertainty tax by a large factor. For a concrete discussion of this, I’ll start by drawing out two such problems, and then give some intuition about some of the design levers available in one of these problems.</p><figure><img alt="Extremely messy, squiggly line on the left end resolving into an uncluttered smooth pen stroke at the right end. My added text says “Network designers must sometimes make peace with uncertainty.”" src="https://cdn-images-1.medium.com/max/1024/1*A7ier35o3jdTW4x0BXX6UA.jpeg" /><figcaption>Image adapted from Damien Newman’s <a href="https://thedesignsquiggle.com/">Design Squiggle</a>. While the Design Squiggle is meant to illustrate the resolution of an uncertain mess into clarity, network designers must often live with the uncertainty.</figcaption></figure><h4>Problem A: CDN server deployment or expansion</h4><p>“Foggy” wants to deploy a new Content Distribution Network (CDN), or add capacity to their existing footprint either by establishing additional sites, or by adding more servers to an existing site. For simplicity, we can assume that the set of possible sites is fixed (e.g., map of existing colocation facilities) and each site is unconstrained in terms of bandwidth, power, server-space, etc. A cost model specifies the cost of adding any number of servers at a particular site. Thus, we can compute the total cost of every candidate deployment assigning a number of servers to each possible location.</p><p>Foggy uses a time horizon, say, one year, for network planning, and has forecasts for the server capacity needed at each location over the next year. The evolution of the ecosystem (e.g., how well content-providers Foggy already partners with do in terms of eyeball-share) will likely cause these capacity forecasts to be a bit off.</p><p>Foggy’s objective is to decide on a server footprint across all sites that will serve its traffic at some target performance level, with minimal performance deterioration if the capacity forecast used for the design is somewhat wrong. In general, Foggy’s deployment should be such that performance degrades gracefully with respect to how wrong the forecast turns out to be.</p><p>What makes the problem interesting is that the deployment that is optimal for the forecast’s expected traffic will in many cases be brittle, and behave poorly for some types of deviations from that forecast. Foggy thus needs to design for the forecast <em>distribution</em> rather than just the expectation. Substantially overprovisioning everywhere will work, of course, but often at a <em>much</em> higher cost than necessary (the uncertainty tax).</p><h4>Problem B: WAN provisioning</h4><p>“Fibby” operates a wide-area network (as an ISP or a private cloud WAN like that of Google/Microsoft/Amazon) for which they want to provision fiber for the foreseeable future. They can build out or lease fiber capacity (and switching infrastructure) on existing or new wide-area network segments for their network. Fibby has forecasts for traffic demand between each pair of end-points on their network, but again these forecasts are inherently imprecise — if Fibby is a transit ISP, their peering agreements with others may evolve or their customers’ needs might shift; if Fibby is a private WAN operator like Google, the changing nature and traction of applications could shift their traffic demands. (Note that the planning horizon here may be several years, as is typical for fiber build-outs and capacity leases.)</p><p>Fibby’s goal is to deploy network capacity such that current and future traffic can be served reliability at high performance, again with minimal deterioration if the forecast for traffic is somewhat incorrect. Similar considerations of uncertainty tax apply to Fibby’s problem as that of Foggy’s.</p><h4>How interesting and new are these problems?</h4><p>At this point, you may rightly point out: “The forecast-reality gap surely occurs everywhere, e.g., a manufacturer deciding how many units of a gadget to order parts for based on forecasts of demand. What’s different about networks, and why haven’t other domains solved this for us?”</p><p>There is a sizable literature on decision-making and optimization under uncertainty, e.g., <a href="https://www.cse.iitb.ac.in/~cs709/notes/readingAssignment/RegretMinimisation.pdf">regret minimization</a>. However, I haven’t seen much work on this for <em>network design, </em>especially more applied work outside theory-focused conferences. (I would be very happy to be pointed to NSDI/SIGCOMM papers that address questions along the lines of the two example problems above.) My expectation is that addressing “network design under uncertainty” will require the development of new methods rather than straightforward application of techniques developed in other contexts.</p><p>The networked nature of networks (I know!) is what makes a difference. Consider the CDN scenario above — if your incorrect forecast for the server capacity needed at one site led you to under-provision that site, the nearby sites <em>might</em> be able to serve the extra traffic, with a latency hit. Thus, decision-making across the network, considering what resources can be moved around with marginal performance penalties, can help protect the operator from sizable forecasting failures. The ability of certain resources to substitute others with some modest/graceful performance degradation can open the door to more aggressive (i.e., cheaper) provisioning at individual sites. Some intuitive examples of using this strategy for Foggy’s problem are as follows:</p><ul><li>Instead of uniformly overprovisioning every CDN site, Foggy could overprovision only a few well-connected, central points (e.g., the Frankfurt/Amsterdam IXPs for Europe), while aggressively provisioning other sites based on their forecast expectation. This could be sufficiently robust in terms of performance, while reducing the costs. In essence, the well-connected points host redundant resources that do a good enough job of making up for shortfalls <em>anywhere </em>in their vicinity.</li><li>Parts of the network that form “communities” separated from the rest of the network by a large latency connection could be provisioned jointly. Each community should ideally have enough overprovisioning internally, such that in most cases, the high-latency backhaul can be avoided.</li><li>Sites that consistently show correlated demand are likely prone to correlated forecasting errors. If several such sites are in the same region, the region may require greater overprovisioning as a whole. Note however, that geographical proximity may not always lead to such correlations, e.g., language borders could lead to proximal locations having very different content requirements and consequently, traffic trends.</li></ul><p>As the above intuitive examples illustrate, a network’s structure and operating context determine how much opportunity there is for network-wide decision making. Thus, I think it would be highly valuable to try to rigorously build methods (beyond simplistic intuitions like the ones I sketched above) for reducing the uncertainty tax in such design problems.</p><h4>Related work</h4><p>I have seen work on verifying probabilistic properties of networks, like the <a href="https://dl.acm.org/doi/10.1007/978-3-662-49498-1_12">ProbNetKat</a> and <a href="https://nsg.ee.ethz.ch/fileadmin/user_upload/publications/netdice-cameraready-sigcomm20.pdf">NetDice</a> efforts, but not much on design-under-uncertainty. The closest pieces of work I’m aware of take a theoretical perspective on a <a href="https://arxiv.org/pdf/2105.02363.pdf">problem similar to Foggy’s</a> above, and a<a href="http://networks.ece.cornell.edu/yingjie/ciss.pdf"> problem similar to but more limited than Fibby’s</a>. There’s also a <a href="https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45385.pdf">Google presentation from 2015</a> about capacity planning that has hints of the problem, but not much detail. It would be great to see more empirical work, informed by real-world data and network structures along these lines. <a href="https://dl.acm.org/doi/pdf/10.1145/3341302.3342069">TeaVaR</a> embodies the <em>type</em> of work I’m missing, but it addresses traffic engineering rather than network design.</p><p>I haven’t been able to thoroughly scour the literature, and am happy to be pointed to other relevant work. It would also be great to see commentary from industry folks who are surely addressing such problems already in <em>some</em> way — is the state-of-the-art in practice just overly cautious overprovisioning?</p><p><strong>Acknowledgments</strong>: Discussions with my former PhD student, <a href="https://ch.linkedin.com/in/vdukic">Vojislav Dukic</a>, and our jointly-supervised former MS student, <a href="https://ch.linkedin.com/in/kenanbesic">Kenan Besic</a>, helped me shape some of my thoughts on this topic.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=891470040d07" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[“NetFlicks”: visualizing network events]]></title>
            <link>https://medium.com/@stub_AS/netflicks-visualizing-network-events-97eb5028371f?source=rss-88a9a115d027------2</link>
            <guid isPermaLink="false">https://medium.com/p/97eb5028371f</guid>
            <category><![CDATA[networking-research]]></category>
            <category><![CDATA[network-visualization]]></category>
            <category><![CDATA[visualization]]></category>
            <category><![CDATA[computer-networking]]></category>
            <category><![CDATA[network-debugging]]></category>
            <dc:creator><![CDATA[Ankit Singla]]></dc:creator>
            <pubDate>Tue, 28 Dec 2021 21:47:43 GMT</pubDate>
            <atom:updated>2021-12-28T21:47:43.541Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This is part of a series of write-ups I’m working on, drawing on my notes and materials over the last roughly 7 years. I’m leaving academia for an exciting industry opportunity after </em><a href="https://twitter.com/stub_AS/status/1469283183132876809"><em>not getting tenure</em></a><em>. The research ideas I’m putting out are open for anyone to build on. The caveat is that my background work on many of these, including a full review of the relevant literature, is incomplete — this is just a snapshot of where I happened to be in my thinking along these lines.</em></p><p>We are still<em>, </em>in 2021, often using the same tools for understanding and debugging networks as we did 30+ years ago, e.g., tcpdumps and traceroutes. Meanwhile, networks have become orders of magnitude larger, more complex, and hard to understand and reason about.</p><p>While there is a rich line of work on making networks more robust, e.g., most recently with network verification and synthesis, and on automated methods of summarizing network logs, I haven’t seen much in the way of tools that aid intuition and understanding by <strong>visualizing </strong>network events and phenomena.</p><p>For instance, instead of examining a tcpdump using Wireshark or (even more painstakingly) the line-by-line text event log, why don’t we have something that shows a movie-like view of the packet sequence, highlighting important or anomalous events?</p><p>The simplest use of this is in networking education, whereby you can show students various events in a packet stream and what parts of the transport algorithm’s state machine they correspond to. For the simplest scenarios, one could imagine a side-by-side view of the network bottleneck’s evolution in terms of buffer state, together with how those events trigger reactions from congestion control logic. (I’m aware of <a href="https://www2.tkn.tu-berlin.de/teaching/rn/animations/gbn_sr/">nice animations</a> of the behavior of two simple transport algorithms, but nothing that uses network logs and data in the way I’m describing here.)</p><p>For more complex use cases, such as the debugging of a complex networked system, it would be great to have a workflow where the following occurs:</p><ul><li>From a network data log or a set of logs, an automated process identifies interesting events. These could be based on several methods, e.g., anomaly detection or triggers encoded by network operator, e.g., “loss exceeds x%”.</li><li>A second step would order events and identify potential causality among the interesting events. This can use a variety of approaches customized to the context in question, e.g., for TCP logs, one could imagine inferring rules like “a decrease in the receiver window could cause a decrease in transmission rate”, “packet loss could cause a decrease in transmission rate”, “packet loss may be preceded by a period of higher RTTs.”, etc.</li><li>Finally, and this where I think prior work is lacking most, a third step visualizes the potential chain(s) of causal events for the network operator as a candidate analysis of the situation. The visualizations can involve several components: the system itself, e.g., an appropriate abstraction of a data center topology showing where loss is occurring and which flows are contributing; the key parts of system state involved, e.g., the buffer(s) at particular switches; and the rules / logic of the algorithmic process that are triggered in response, e.g., congestion avoidance mechanisms; etc.</li></ul><p>Each of the three steps mentioned above are far from trivial, but the thing is: if this “NetFlicks” system can build a one-minute movie showing the operator a plausible evolution of their system, it doesn’t have to always get it right. If it gets it wrong, the operator can always go back to whatever methods they’re already reliant on. But if NetFlicks gets it right, the operator might learn the needed insights in a minute rather than hours or days of painstaking analysis. The vast gap in the usefulness when correct versus the annoyance when mistaken seems to set a low bar for the first tools along these lines. How cool would it be to have a plausible diagnosis of a situation that can be understood in a minute, even if it’s only right (or nearly-right), say, 20–30% of the time?</p><p>Of course, such a tool will need to be customized to every context, with different tooling for a satellite network than for a data center network, and different tooling to debug routing vs. transport issues, with things that involve both being even more complex.</p><p>To make the above discussion more concrete, consider a hypothetical NetFlicks extension of the <a href="https://github.com/snkas/hypatia">Hypatia framework for visualizing satellite networks</a> which might find the following set of events and causes:</p><ul><li>[Anomalous event] Several connections are experiencing very low throughput, even though the links they use are under-utilized.</li><li>[Causality-A] The delay-based congestion control logic has throttled the sending rate.</li><li>[Causality-B] Queuing delays are fluctuating.</li><li>[Causality-C] The end-end path sometimes elongates due to satellite motion (increasing the delay even though there’s no congestion).</li></ul><p>Now, across a set of traces, it should be possible to identify that C leads to B leads to A leads to the low throughput event, and to display the following in a “scene”:</p><ul><li>An example path change, highlighting the elongation of the path, for instance like the change in the below graphic, but just continuous instead of looped.</li><li>Side-by-side, the corresponding throughput evolution over time.</li><li>Side-by-side, the part of the congestion control logic that is triggered at the point(s) throughput drops or increases substantially.</li></ul><p>I was lazy here, but I very well could have put together the desired movie for such a scenario <strong>manually</strong> if I wanted, using Hypatia. But the goal of a project along these lines should be to help uncover and visualize such insights in as automated a fashion as possible. Of course, a variety of pieces will need to be manually provided up front, e.g., ways to visualize the system at any time, encoding what types of events could be interesting, what metrics might be interesting to monitor the temporal evolution of, etc.</p><figure><img alt="A gif-image  shows a path over an LEO satellite network changing in length." src="https://cdn-images-1.medium.com/max/1024/1*bt-c3-JA0KUTVhdPjTtQ3Q.gif" /><figcaption>The path elongation (worth several tens of seconds here) can cause delay changes that cause delay-based congestion control to mistake the delay increase for congestion and throttle throughput.</figcaption></figure><p>Ideally, NetFlicks runs through “normal” operation speedily, and slows down to highlight the part crucial to understanding what’s going wrong during the throughput deterioration event. Sort of like slow-motion enhancements used in action movies and the below adorable cat-action (which I most definitely did not include only to capture your attention):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/728/1*mB6qe-A3FMEVvvmACAlP8Q.gif" /><figcaption>From <a href="https://youtu.be/YHdOh90kGHQ?t=48">“The Slow Mo Guys” on YouTube</a>.</figcaption></figure><p>The same type of visual reasoning, with different base visualizations of the system and different events and control logic encoded, should be applicable for a variety of networked contexts, e.g., for a data center environment, or a Google/Microsoft-style inter-datacenter WAN.</p><p>In my view, this is a reasonably concrete project vision that could produce HotNets-like work (showing very basic functionality) within 4-ish months of work, with a healthy potential for leading to a series of papers and software as part of a solid dissertation, and (very optimistically) even the foundation of a startup around this type of visual network intelligence that complements tools sophisticated network operators use.</p><p>I am looking forward to finding out (hopefully soon) how some of the most complex and best-managed networks today my are managing network debugging and building understanding and intuition, but meanwhile, if anyone is willing to share anything in the state of the art along the lines I’ve sketched, please do. I’m reasonably sure that there are reams of related literature I could have cited that provides various pieces that would be of use for building NetFlicks, and if you want to leave pointers in comments, that’s more than welcome. If I find the time and inclination, I can try to update this writeup with the most salient pointers.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=97eb5028371f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Grant applications]]></title>
            <link>https://medium.com/@stub_AS/grant-applications-da632d7d897f?source=rss-88a9a115d027------2</link>
            <guid isPermaLink="false">https://medium.com/p/da632d7d897f</guid>
            <category><![CDATA[academia]]></category>
            <category><![CDATA[grant-research]]></category>
            <category><![CDATA[grants-money]]></category>
            <dc:creator><![CDATA[Ankit Singla]]></dc:creator>
            <pubDate>Mon, 20 Dec 2021 22:00:53 GMT</pubDate>
            <atom:updated>2021-12-20T22:00:53.453Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This is part of a series of write-ups I’m working on, drawing on my notes and materials over the last roughly 7 years. I’m leaving academia for an exciting industry opportunity after </em><a href="https://twitter.com/stub_AS/status/1469283183132876809"><em>not getting tenure</em></a><em>.</em></p><p>In the 5 years from 2016-2020 when I was actively applying for grants, I raised more than enough money to support my group, which at its peak comprised 7 PhD students. (A good chunk of my resources were ETH provided, which included substantial base funding, startup money, and matching funds for some grants.) Funding was unproblematic, with very limited time spent on it, and with easy prospects for raising more money if I needed it. But especially in later years, once I had overcome the problem of finding the first few great students to get my group going, I was constrained by my own bandwidth rather than funding — I just didn’t think that I could support a larger group than 7-8 students without either compromising on the quality of supervision each student received, or delegating substantial parts of supervision to post-docs, something I was quite reluctant to do.</p><p>A majority (roughly 60%) of my external non-ETH funds came in the form of industry gifts, some of which were received on the backs of personal engagement, with only short emails describing the concrete research ideas and the intended scope of work around them, instead of extensive proposals. In one instance, I benefited from industry connections my senior colleagues had nurtured. In every instance, I retained full control over the research scope. Even for the industry money, it wasn’t “contracted research” — the only commitments I made in return for the money were early access to research results and continued engagement via yearly presentations and meetings.</p><p>For the competitive proposals across both industry (e.g., Google’s Faculty Research Award) and academia (e.g., the Swiss National Science Foundation’s Projects), I found that they were even less predictable than research paper outcomes: some proposals, which I spent months writing, and for which I received positive feedback from senior colleagues, were met with crushing rejection, while others that I rushed in within days were accepted. For almost all grants I got rejected, the results of the work were eventually published at top competitive venues, with some even winning awards. I can only speculate why the grant process seems so fickle, but one potential reason is just high reviewing randomness — grant reviews were typically poorer in quality than those at top publication venues, and for a couple of high-value applications, just one reviewer’s confidently-wrong opinion killed them.</p><figure><img alt="The Hildebrandt brothers’ Smaug, guarding its treasure with fire. The image is modified to label Smaug as “Reviewer 2&quot;, and the object its fire and fury as “Grant app”." src="https://cdn-images-1.medium.com/max/605/1*6sCcyFb1WgST1wyOWaKc3g.png" /><figcaption>“Reviewer 2” crushes many spirits before a paper is ever submitted. (The image is a modified version of the <a href="https://scifinet.net/greg-and-tim-hildebrandt/">Brothers Hildebrandt illustration of Smaug</a>.)</figcaption></figure><p>I’m linking below my significant grant applications, and given that I do not intend to pursue further work on these threads, the ideas here that haven’t appeared already in my papers are available to anyone who wants to run with them. Note: the “results” in these applications were obtained at preliminary stages of work, and thus are not definitive or citation-strength evidence.</p><ul><li><a href="https://drive.google.com/file/d/197KfIfw6cMViyvOHGCznHhUMss_fcztp/view?usp=sharing">Swiss National Science Foundation (SNSF) Project, 2016</a></li><li><a href="https://drive.google.com/file/d/1vrMFddYvHje0ABHDvgkLRV-Ty4J55l_5/view?usp=sharing">Swiss National Science Foundation (SNSF) Project, 2017</a></li><li><a href="https://drive.google.com/file/d/1pGTrIOmezaJnnqIcDsHxN0mFrNlZ7slv/view?usp=sharing">Swiss National Science Foundation (SNSF) Project, 2018</a></li><li><a href="https://drive.google.com/file/d/1fdZb3VECrqWoRdvdZBDjybTlb36DSURo/view?usp=sharing">Google Faculty Research Award, 2016</a></li><li><a href="https://drive.google.com/file/d/1GQJ5wJ1CR1HblAHVs39dMIvUWAcYFeXc/view?usp=sharing">Google Faculty Research Award, 2017</a></li><li><a href="https://drive.google.com/file/d/1Zak2PFRyLFlbx4CSDZGVnuiHWl7cU6Y_/view?usp=sharing">Google Faculty Research Award, 2018</a></li><li>ERC Starting Grant, 2018 — This was on <a href="https://cspeed.net/">cSpeed</a>. I’m leaving out the documents for this, because a part of the grant touches on ideas I started working on with some collaborators who continue to develop them.</li><li>ERC Starting Grant, 2020 — the <a href="https://drive.google.com/file/d/1EQoBCQBwRz146vUiCCPztMg5JRFGezVx/view?usp=sharing">short-form</a> and <a href="https://drive.google.com/file/d/1lOAzgd6DdYRgLdhskQRv3RG68gMBkduY/view?usp=sharing">long-form</a> documents</li></ul><p>I want to draw out only a few notes from my experience, which is fairly limited, both in the timespan, and the geography / funding bodies:</p><ul><li>Write and submit workshop papers on ideas you want to develop into grants. This helps not only get early feedback from your community, but also gives you great raw materials for grants. For collaborative papers, you should be clear about who owns what for grant writing.</li><li>Develop writing structures that clearly spell out the research questions after the contextual ground is prepared. I like the structure in my SNSF apps, although the condensed version in the ERC also seems fine to me.</li><li>Industry money comes in different flavors. Some, like the Google Faculty Research Awards, is as or more competitive as more traditional funding bodies, even though the grant money involved is small. The big money in industry funding, as far as I can tell, is in significant part, gated on deep, long-running personal ties with industry contacts.</li><li>Get the table-setting, formatting, bureaucracy, letters of support, etc. out of the way well before the application deadline. Even getting the required formatting right at crunch time can be a painful time sink.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=da632d7d897f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[An account of an academic job search]]></title>
            <link>https://medium.com/@stub_AS/an-account-of-an-academic-job-search-ad8df6eb3031?source=rss-88a9a115d027------2</link>
            <guid isPermaLink="false">https://medium.com/p/ad8df6eb3031</guid>
            <category><![CDATA[the-academic-job-market]]></category>
            <category><![CDATA[academic-jobs]]></category>
            <category><![CDATA[faculty-jobs]]></category>
            <dc:creator><![CDATA[Ankit Singla]]></dc:creator>
            <pubDate>Tue, 14 Dec 2021 11:35:20 GMT</pubDate>
            <atom:updated>2021-12-14T11:35:20.219Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This is part of a series of write-ups I’m working on, drawing on my notes and materials over the last roughly 7 years. I’m leaving academia for an exciting industry opportunity after </em><a href="https://twitter.com/stub_AS/status/1469283183132876809"><em>not getting tenure</em></a><em>.</em></p><p>I’ll try to give as complete a view of my search for a tenure-track “Assistant Professor in Computer Science” position, interviewing in Spring 2015, as reasonably possible. I’m not going to make an effort to try and distill too many lessons from this, but I hope folks who intend to pursue similar objectives, as well as the departments hiring them, find some useful tidbits in here.</p><h4>Application materials</h4><p>There are many other great examples you can find on the Web. For PhD candidates approaching their job market hurrah in a year or two, I recommend looking at folks in the market right now, especially in their own and adjacent areas, and saving their materials (typically on their Web pages) to refer back to later. Mine are here:</p><ul><li><a href="https://drive.google.com/file/d/18A7Lz6d33Mhho6oHLTY607DVMxAeEdW9/view?usp=sharing">CV</a></li><li><a href="https://drive.google.com/file/d/1NhbmaoltswF54KH1Gch5f7Qquff4mfRR/view?usp=sharing">Research statement</a></li><li><a href="https://drive.google.com/file/d/1nmFTyod8Hy2qEI6K1L_LZHhgMPoVSZmx/view?usp=sharing">Teaching statement</a></li></ul><h4>My interview plan for one-on-one conversations</h4><p>At the start of the search, I wrote <a href="https://docs.google.com/document/d/16MEb1zVlkJ3U5lxRoEJ5dEo3z8IrnJ9WId4xWikYPCs/edit?usp=sharing">a document</a> containing a “recipe” to help myself steer and manage the one-on-one conversations with individual faculty over campus interviews. It contains the following:</p><ul><li>A generic pitch about myself, for those who didn’t attend my talk. I had a 1-minute version and outlines for a detailed 5–10 minute version.</li><li>A list of research ideas that I had come up with and could potentially pursue with my future students.</li><li>For each campus interview, once I got the schedule, I would take notes on each person I would be meeting, and what I could talk to them about.</li><li>I anticipated questions people would ask me, and also listed questions I wanted to ask them.</li></ul><h4>My job talk</h4><p>UC Irvine posted my interview talk’s recording to <a href="https://www.youtube.com/watch?v=nEU34Xl0nYk">YouTube</a>, so it is available for everyone to see.</p><h4>My diary, with a full play-by-play</h4><p>This is a <a href="https://drive.google.com/file/d/1K7kku9qfoD2MRpoihPu-ny2qNImSVT_f/view?usp=sharing">play-by-play record</a> of my academic job search, exactly as it went down, with no changes to text from the 2015 account. It contains everything from the point of preparing application materials, bugging letter writers, the arrival of phone and campus interviews, how they went, what impressions I formed, some odd/weird things people said or did, etc. Tiresome to read at length, but I have added some comments with the benefit of hindsight at some points.</p><p>I made redactions of identifying information for the places and people I interviewed with, and precise dates I interviewed at those places on. I didn’t feel comfortable putting folks on blast for how they did things 6 years after the fact — I don’t know how things changed in the intervening years, and hopefully, there were improvements across the board. I think people familiar with how interviews work at various top departments might have good guesses for some places, but so be it.</p><p>It’s always a bit cringe-inducing to reflect on one’s past self, and it was no different for me, especially the casual manner in which I cast broad judgments on people and places. I ask your forgiveness for my 2015-self for my obtuseness on some things :)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ad8df6eb3031" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Research problems in satellite-based broadband, à la SpaceX Starlink]]></title>
            <link>https://medium.com/@stub_AS/research-problems-in-satellite-based-broadband-%C3%A0-la-spacex-starlink-6f5dfe6d2756?source=rss-88a9a115d027------2</link>
            <guid isPermaLink="false">https://medium.com/p/6f5dfe6d2756</guid>
            <category><![CDATA[research-idea]]></category>
            <category><![CDATA[spacex]]></category>
            <category><![CDATA[leo-network]]></category>
            <category><![CDATA[starlink]]></category>
            <category><![CDATA[satellite-networking]]></category>
            <dc:creator><![CDATA[Ankit Singla]]></dc:creator>
            <pubDate>Fri, 10 Dec 2021 11:56:07 GMT</pubDate>
            <atom:updated>2021-12-10T11:56:07.638Z</atom:updated>
            <content:encoded><![CDATA[<p>I’ll first give brief context on the recent industry developments and why they matter. Then I’ll outline several research directions that I believe will lead to both top-tier publications and long-term industry impact. Lastly, to help you get started, I’ll point you to more in-depth readings and tools for research.</p><h4>Thousands of satellites providing low-latency broadband Internet</h4><figure><img alt="An illustration of orbital shell, orbital planes, and satellites planned to comprise part of the SpaceX Starlink constellation." src="https://cdn-images-1.medium.com/max/743/1*uV8WRvbpk5hZ3qkB2l5NgA.png" /><figcaption>Five of Starlink’s planned orbital shells coded into different colors. Some shells, e.g., the dense red and green ones, do not cover the Earth’s poles.</figcaption></figure><p>As of December 2021, Starlink comprises 1,732 working satellites, making it the <a href="https://en.wikipedia.org/wiki/Satellite_constellation#List_of_satellite_constellations">largest constellation</a> that has ever been deployed<em> by far</em>. Altogether between Starlink, OneWeb, Amazon’s Kuiper, Telesat, and many others, more than 10,000 satellites may be deployed by 2030. These constellations will all operate in low Earth orbit (LEO), a few hundred to at most 2,000 km above the surface, and provide low-latency broadband across the globe. Note that with LEO satellites (unlike for satellites in Geostationary orbit, GEO), ground-to-satellite-to-ground latencies can, in theory, be well under 10 ms. For communication between endpoints separated by a few thousand kilometers or more, LEO networks (with inter-satellite laser links deployed) can provide <a href="https://dl.acm.org/doi/10.1145/3286062.3286079">lower latency than even terrestrial fiber</a>.</p><h4>Why is LEO broadband a big deal?</h4><p>Even in the most pessimistic scenario, LEO networks will soon serve millions of consumers. This is just the low-hanging fruit, because satellite-based services that are clearly inferior (much higher latency using GEO) are already serving around <a href="https://en.wikipedia.org/wiki/Viasat_Internet">2 million</a> <a href="https://en.wikipedia.org/wiki/Hughes_Network_Systems">customers</a> in the Americas alone. A more optimistic outlook may suggest hundreds of millions of users globally within a decade, with many served indirectly, such as by LEO satellites providing backhaul connectivity for terrestrial cellular towers in rural/remote areas. Further, LEO connectivity has myriad applications beyond consumer broadband, including disaster management, inflight connectivity, and shipping, offshore, and other remote-site industries. Even cloud operators like Google and Microsoft, with state-of-the-art terrestrial networks, have <a href="https://www.prnewswire.com/news-releases/google-cloud-and-spacexs-starlink-to-deliver-secure-global-connectivity-301290801.html">inked</a> <a href="https://news.microsoft.com/transform/azure-space-partners-bring-deep-expertise-to-new-venture/">partnerships</a> with Starlink to benefit from its global low-latency reach.</p><p>The below graphic summarizes the satellite network design space, capturing (coarsely) the relationship between some of the design options and their outcomes along the axes of latency and capacity/market-size. (I’m conflating the latter two, which are aligned, but not the same thing.)</p><figure><img alt="Plot with two axes capturing design parameters for satellite networks. X-axis captures network capacity and market size, going from niches to “the next billion Internet users”. Y-axis describes network latency. A few LEO satellites can serve low-latency traffic for niche apps. A few GEO satellites can serve larger bandwidth, still for limited users, but at high latency. Massive-scale LEO can do both low-latency and high throughput for a large user population." src="https://cdn-images-1.medium.com/max/702/1*7DJyC7QsWdYKKPA-yd71lw.png" /><figcaption>Massive-scale LEO can provide both low-latency and high throughput for a large user population, depending on how many satellites are used in the system and the per-satellite bandwidth.</figcaption></figure><h3>Open research challenges and opportunities</h3><p>At my <a href="https://youtu.be/-kxgBI1osl0">opening talk for LEOCONN 2021</a>, the first workshop on satellite mega-constellations, I outlined 20 research challenges in this area, most of which I believe can and will be developed into top-tier research publications, and will inform and influence future industry activity. I will only touch on three broad sets of questions in this writeup, pointing interested readers to the talk’s recording (<a href="https://youtu.be/-kxgBI1osl0">https://youtu.be/-kxgBI1osl0</a>) for the rest.</p><h4>Opportunity 1: Mixed-orbital constellation design</h4><p>The easiest-to-explain research opportunity uses the context I set up just above in describing some of the design options for satellite networks. The idea is simple: combine the strengths of LEO-only and MEO-only networks. (MEO networks operate in <a href="https://en.wikipedia.org/wiki/Medium_Earth_orbit">medium Earth orbit</a>.)</p><p>LEO networks can provide low-latency, but being closer to the ground means that each satellite’s coverage cone is small, and you need many hundreds if not thousands of satellites to provide coverage (at great expense). MEO networks (at certain orbital heights) can provide global coverage with a few tens of satellites while still providing sub-50 ms latencies. Thus, one could imagine building a system with (say) a hundred LEO satellites and a few tens of beefier MEO satellites, such that low-latency traffic (typically also low in volume) uses LEO connectivity, while throughput-focused traffic uses MEO. This would likely result in a substantially cheaper network, albeit requiring careful partitioning of traffic across the mixed orbital design. I’ve positioned this proposal in the design space schematic below. The challenges include, for instance, figuring out how best to combine different types of orbits for specific targeted traffic mixes, how to partition traffic in a reasonably transparent way (hiding complexity without violating neutrality expectations), and how the optimal design might evolve as satellite launch costs shift and externalities like space-congestion and night-sky-pollution are priced.</p><figure><img alt="On the same axes/plot as the previous image, a new design choice is added, showing a combination of LEO satellites for low-latency traffic with a few high-throughput MEO satellites for dealing with bulk traffic." src="https://cdn-images-1.medium.com/max/643/1*yClFAa-aSaMGeEfvSh4inw.png" /><figcaption>A system combining a few hundred LEO satellites and a few tens of MEO satellites could be less expensive than mega-constellations and provide both low-latency and high throughput for different traffic classes.</figcaption></figure><h4>Opportunity 2: Congestion control and traffic engineering</h4><p>Our <a href="https://bdebopam.github.io/papers/imc2020-hypatia.pdf">measurement and analysis work</a> showed that as LEO satellites zoom around in Space at 27,000 km/h, the structure of paths and end-end latencies evolve over time. These changes cause packet reordering (as paths sometimes shorten, and earlier-in-sequence packets arrive after later-in-sequence packets), which congestion control can misinterpret as loss. Loss is also an undesirable primary signal of congestion for an entirely separate reason: operating with full or substantially occupied buffers would negate the low-latency promise of these networks. However, the other potential congestion signal — increasing delay — is also noisy in this case, as paths naturally elongate and shorten independent of congestion.</p><p>It is an open problem how to design congestion control for such conditions where both loss and delay are noisier and more unreliable than in typical terrestrial conditions. Of course, if one were to “subtract” the delay changes caused by (predictable) satellite motion, the problem seems easy, but the issue is that this requires a break in abstraction: now the endpoints have to be network aware to be able to perform this “subtraction”. In general, an end-end path may traverse both LEO segments and ground segments, and one should not expect endpoints to care about the network-layer path in making their transport-layer decisions.</p><p>Another potential solution involves transport proxies that terminate endpoint connections and can, by virtue of sitting at the LEO network edge as middleboxes, appropriately account for satellite motion in making congestion control decisions. This would be a break from the end-end principle, and comes with costs: if the proxy dies, connection state would be lost, or worse, could be corrupted (e.g., the proxy ACKs some data to the sender and then dies, without the receiver seeing the data). Equally worrisome, this model would also not work with end-end encrypted QUIC transport.</p><p>Finally, the utilization of certain network links can change dramatically as they end up on a route between traffic hotspots, or move off of such routes. (A small visualization of this is in the example below; details are in <a href="https://bdebopam.github.io/papers/imc2020-hypatia.pdf">our paper</a>.) Such changes imply that congestion control needs to converge fast in the presence of shifting bottlenecks and the extent of congestion on them. Some of this <em>could</em> potentially be tackled at the network layer, with careful traffic engineering to smoothen out the impact of upcoming path structure changes by gradual shifts in traffic across multiple routes.</p><figure><img alt="A gif-image showing visually how utilizations of certain links change over time on an end-end path because traffic shifts." src="https://cdn-images-1.medium.com/max/1024/1*YO12nnRNoCEsOQxtN3XiDQ.gif" /><figcaption>The utilization of different links on an end-end path in an LEO network is color coded here, with thin/green being low utilization, and warmer colors or thicker lines indicating higher link utilization. Notice the link on the far left and the links in the center-right part of the image — their utilization changes across the snapshots.</figcaption></figure><p>Thus, I expect there to be interesting followup work on rigorously evaluating numerous current congestion control proposals in challenging LEO settings, as well as in designing new congestion control and traffic engineering schemes that address the issues outlined above.</p><h4>Opportunity 3: Integration with cloud and content networks</h4><p>I noted above the Google and Microsoft partnerships with Starlink. If you’ve made it this far into the writeup, you are likely familiar with the bespoke WAN designs of these operators, <a href="https://dl.acm.org/doi/10.1145/3230543.3230545">B4</a> and <a href="https://conferences.sigcomm.org/sigcomm/2013/papers/sigcomm/p15.pdf">SWAN</a> respectively. These networks use software-defined networking techniques to manage global inter-datacenter connectivity. The Starlink partnerships open up a new and interesting set of research questions:</p><ul><li>What customizations do the B4/SWAN networks need to incorporate the dramatically different connectivity LEO networks provide? Are the existing solutions already capable of taking advantage of LEO links with link characteristics that are very different from terrestrial fiber links in terms of packet reordering, latency variation, bit error rates, and random loss?</li><li>Where both terrestrial and LEO connectivity are available, and the LEO link offers lower latency (for long distances, as noted above), how to best use limited LEO capacity?</li><li>Is there value in sophisticated operators like Google and Microsoft getting the raw channel, minus forward error correction, from the LEO operator? One could, for instance, imagine the cloud operators wanting to customize the degree of FEC to arrive at the best operational tradeoff with latency. (Analogously, some customers of terrestrial microwave networking, such as high-frequency traders, typically get an FEC-free channel, as they place a premium on latency.)</li><li>If the cloud operator wanted to resell LEO connectivity as a premium enhanced turbo nitro networking service to their customers, how would they define SLOs for such time-variable connectivity?</li></ul><p>Lastly, beyond the announced partnerships by Google and Microsoft, can content distribution networks also benefit from the low-latency connectivity offered by LEO networks? If so, how should a CDN architecture incorporate LEO connectivity and potentially, <a href="https://bdebopam.github.io/papers/hotnets20-inorbit.pdf">on-satellite storage and compute</a>?</p><h3>Resources to help you get started with LEO research</h3><ul><li>For a more fleshed out read on the potential and challenges of LEO networks, I recommend our <a href="https://bdebopam.github.io/papers/hotnets18_space_race.pdf">HotNets 2018 work</a>. (2 hours)</li><li>For learning about the nuts and bolts of an LEO network — what parameters describes such a network’s structure, how connectivity works — I suggest section 2.1 in<a href="https://bdebopam.github.io/papers/conext19_LEO_topology.pdf"> our CoNEXT 2019 paper</a>. (1 hour)</li><li>If you are searching for interesting research problems to explore, check out my <a href="https://youtu.be/-kxgBI1osl0">opening talk at LEOCONN 2021</a>, where I touch upon 20 open research challenges in satellite networking. (30 min)</li><li>We spent substantial effort in building simulation tools that bake in many of the abstractions that would be new to networking researchers, such that the bar for entry into this new research area is as low as possible, with different frameworks for different types of work at the <a href="https://github.com/satnetwork/satnetwork.github.io">graph/topology level</a>, for <a href="https://github.com/giacgiuliari/icarus-framework">routing</a>, and for <a href="https://github.com/snkas/hypatia">full-fledged packet-level simulations</a>.</li></ul><p>I am grateful to all my co-authors and collaborators on the above listed papers and other work in this area for shaping my thoughts on this exciting space. <a href="https://bdebopam.github.io/">Debopam Bhattacherjee</a> deserves a special thanks — as is often the case with excellent PhD advisees, I learnt as much or more by working with him as he did by working with me :-)</p><h3>Closing notes</h3><p>In my view, mega-scale LEO networks are “one giant leap” in Internet infrastructure, and could very well improve a billion lives (or more!) in the long run. There is a huge amount of opportunity for innovation at every layer of the network stack, and for every stakeholder in the networking ecosystem. I only sketched out a few research ideas here, and many more are presented in <a href="https://youtu.be/-kxgBI1osl0">my LEOCONN talk</a>. I’m also certain that there are entire sub-areas that my writeup is missing, e.g., there must be many challenges just in the ground-to-satellite radio segment. I thus hope to see an explosion of activity on this front at top-tier networking research venues.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6f5dfe6d2756" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to write a rebuttal]]></title>
            <link>https://medium.com/@stub_AS/how-to-write-a-rebuttal-31e93c4f4b3f?source=rss-88a9a115d027------2</link>
            <guid isPermaLink="false">https://medium.com/p/31e93c4f4b3f</guid>
            <category><![CDATA[rebuttal]]></category>
            <category><![CDATA[academic-writing]]></category>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[conference-review]]></category>
            <dc:creator><![CDATA[Ankit Singla]]></dc:creator>
            <pubDate>Wed, 17 Mar 2021 19:46:16 GMT</pubDate>
            <atom:updated>2021-03-17T19:52:59.490Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Note:</strong> I wrote this for my <a href="https://nd.ethz.ch/">own research group</a>. Our rebuttals go to conferences in the Computer Systems research area. Typically, the rebuttals ask for a very short response of a few hundred words. I also highly recommend <a href="http://matt.might.net/articles/peer-review-rebuttals/">Matt Might’s advice</a> on the same topic, although I think it’s more geared towards longer-form rebuttals, where there’s an opportunity to respond to all comments.</p><h4><strong>Preliminaries</strong></h4><ul><li>Remember that reviewers probably read your paper weeks before you’re reading their comments. By the time they see your rebuttal, they will have forgotten a lot of the paper, and their own comments. This is why you need <strong>structure </strong>and enough detail in the rebuttal. You have to pick and choose what to respond to.</li><li>The chance that your rebuttal will move decisions is still small, but especially if you got a mix of scores (some positive, some negative), there’s still value in writing a rebuttal — moving just one reviewer your way can change the outcome. You’ve spent maybe a year (more?) on the work, so spend a few days on the rebuttal.</li><li>Follow the conference’s instructions, e.g., typically, rebuttals are asked to try to clear up misunderstandings, answer direct reviewer questions, etc. Promising additional work will usually not help, e.g., saying “We will compare our system against SYSY, as suggested by reviewer X.” This is nice, but reviewer X doesn’t know whether you’ll follow through or not, and they probably won’t change their mind due to this promise. (Note that this stems from the one-shot nature of rebuttals in conferences where I publish; for venues where there’s a back and forth between reviewers and authors, making promises can help.) Instead, you could (maybe) argue, e.g., “We could not compare with SYSY because its code is not available. Qualitatively however, SYSY’s approach is not applicable in a, b, c scenarios, while our solution is more general. This is because …”</li><li>Don’t hesitate to acknowledge that the reviewers are right about something, and the paper should have addressed it more carefully. Reviewers are subject matter experts; you’re not going to bluster your way to acceptance successfully. Respond to the critique and mention that you will add a response along the same lines to section X.</li></ul><h4><strong>The process of writing a rebuttal</strong></h4><ul><li>Try to group different comments into umbrella topics, and organize the rebuttal by topic, not by reviewer — reviewers will often want to see responses to each others comments, so a topic organization is likely to be more suitable.</li><li>Topics should ideally not be too coarse, e.g., instead of “evaluation”, use “scale of testbed experiments”.</li><li>Don’t respond to every comment; try and figure out which comments were “deal breakers” for which reviewers. If space permits, make a separate catch-all topic in addition to the topics above, e.g., “other specific queries”, where you can address some standalone, brief comments / questions from reviewers.</li><li>For each topic, label it with which reviewers’ concerns are addressed in it, e.g., “Scale of testbed experiments (Reviewer A, C, E)”.</li></ul><h4>A step-by-step procedure drawing on the the above points</h4><ol><li>Copy all reviews, as is, into a document</li><li>Write a review summary, by going through each review, and picking out comments. You can already leave out anything you don’t perceive as important to address.</li><li>Ideally, at this point, you should have a good sense of a handful of concerns that are critical, and fall into the “must address” category in your rebuttal.</li><li>Now frame topics that seem to cover most comments, or which multiple reviewers care about. The review summary should be what guides this framing of topics.</li><li>Pick colors for each topic, and paint / highlight all the full-review text which is covered by each topic by its corresponding color — this helps you later see what things got left out. This makes sure that while you used the review summary and your impressions to guide the top-level structure of your rebuttal, you aren’t missing any crucial detail.</li><li>Write out the response for each topic. You may have to go back to the colored text for this topic in each review to see if you addressed all the different reviewer questions / perspectives on this topic.</li><li>Now go over what remains unmarked / uncovered in the full reviews: are any key questions unanswered? Put responses for these into the catch-all topic. Here, you could use each bullet-point for one query, and label it with the reviewer to whom it is addressed.</li></ol><h4><strong>Fine-tuning</strong></h4><ul><li>Try to rephrase things in their simplest form; same kind of refinement process as for writing papers applies, except perhaps even more geared towards being concise.</li><li>If you’re over space, cut out what you perceive to be the least important concerns.</li><li>Avoid pointing to sections in the paper too much; you don’t want to risk sounding like “It’s in section 5.6, you blind idiot.” (even if that’s how you felt).</li><li>The ordering of the topics should be governed by two things: (a) perceived importance of the reviewer concerns on the topic; and (b) the strength of your argument rebutting those concerns. So, even if topic-a is somewhat more important (e.g., more reviewers commented on it, or the more negative reviewers picked on it) but topic-b plays to your strength (and was still quite high up in importance), then you may want to put topic-b first in the ordering. This will only work if topic-b wasn’t too far behind topic-a in importance.</li><li>Many CS conferences use HotCRP for reviews. The site supports formatting to a limited extent, e.g., bulleted lists work, and text can be italicized or made bold. Use this markup, e.g., bold for your topic categories or/and your reviewer call-outs.</li><li>Note that some reviewers won’t bother going to the HotCRP site to read your rebuttal, instead just seeing it in their email. (By default, HotCRP sends each reviewer a notification with your rebuttal text.) This will be plain-text, so make sure your markup isn’t too annoying in plain-text form. You can also use formatting like all-caps for topic categories or/and your reviewer call-outs “SCALE OF TESTBED EXPERIMENTS (REVIEWER A, C, E)”, which will work even in plain-text.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=31e93c4f4b3f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>