<?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 Derek Burke on Medium]]></title>
        <description><![CDATA[Stories by Derek Burke on Medium]]></description>
        <link>https://medium.com/@derekbburke?source=rss-9c1626f353e4------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Cqsqbe9YbDun54BTi-1cSw.jpeg</url>
            <title>Stories by Derek Burke on Medium</title>
            <link>https://medium.com/@derekbburke?source=rss-9c1626f353e4------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 02:59:53 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@derekbburke/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[The 4 Worst Mistakes That Slow Your GPU Computing — and How to Avoid Them]]></title>
            <link>https://derekbburke.medium.com/the-4-worst-mistakes-that-slow-your-gpu-computing-and-how-to-avoid-them-fc9fba21e4d2?source=rss-9c1626f353e4------2</link>
            <guid isPermaLink="false">https://medium.com/p/fc9fba21e4d2</guid>
            <category><![CDATA[gpu]]></category>
            <category><![CDATA[data]]></category>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Derek Burke]]></dc:creator>
            <pubDate>Sat, 01 May 2021 00:00:00 GMT</pubDate>
            <atom:updated>2021-05-01T15:37:41.477Z</atom:updated>
            <content:encoded><![CDATA[<h3>The 4 Worst Mistakes That Slow Your GPU Computing — and How to Avoid Them</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lgX0HGq2GwYhuZB1" /></figure><p>GPU-accelerated compute nodes have up to 50x the performance of CPU-only systems.</p><p>Their ‘performance density’ is huge!</p><p>For example, in 2008, the world’s first petaflop supercomputer (Roadrunner at Los Almos National Labs) had 130,464 CPU cores and filled 12,432 rack units. By contrast, NVIDIA’s latest DGX-A100 compute server, has 5x the performance in only six rack units.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TWz-ytoTn2JnsHSB9cq_Fw.png" /></figure><p>The latest GPU compute nodes are not only ‘data hungry’, they are the competitive eating champions of the compute world. That’s why they have multiple 100Gb or 200Gb network connections to keep them busy with data.</p><p>Why, then, would you starve them of data?</p><p>You wouldn’t.</p><p>At least not on purpose.</p><p>To avoid starving your GPUs, here are the 4 worst mistakes that slow your GPU computing — and how to avoid them.</p><h3>1. Forgetting About Data Storage Performance per TB</h3><p>Datasets in HPC and AI can be huge. They are typically stored in network-attached storage systems where data is distributed across multiple drives and presented to the compute systems as a single shared namespace. The ‘aggregate performance’ of the storage system dictates how fast it reads and writes data.</p><p>Let’s look at actual performance data from two all-flash storage solutions to see how this can affect the performance of your data.</p><p>Solution 1 has 500TB of usable capacity and has 2:1 compression with an effective capacity of 1PB. It’s aggregate read performance is 40GB/sec and write performance is 5GB/sec. For every TB of data the read performance is 40MB/sec and write performance is 5MB/sec.</p><p>Solution 2 has 1PB of usable capacity. It’s aggregate read performance is 380GB/sec and write performance is 120GB/sec. For every TB of data the read performance is 380MB/sec and it’s write performance is 120MB/sec.</p><p>Solution 2 has 9.5x faster read performance and 24x faster write performance per TB of data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vYH64nUo6pScRhwHRFAQew.png" /><figcaption>Solution 2 has 9.5x faster read performance and 24x faster write performance per TB of data</figcaption></figure><p>Why on earth would you buy solution 1?</p><p>To be frank, it’s smoke and mirrors. Solution 1 is pitched as <em>“the performance of flash for the price of spinning disk”</em>. But, per TB, it’s slower than some disk-based solutions. During the sales process you will be presented/distracted by an attractive ‘ten year total-cost-of-ownership comparison’. That’s <strong>ten years of slow data</strong>, during which, your applications and clients will become more data hungry. Don’t make that mistake!</p><h3>2. Copying Data into Compute Nodes Before Processing</h3><p>GPU-compute nodes can have significant amounts of internal RAM and disk space. What’s faster than having your data there? You can copy data into your compute nodes and then process it. Then you don’t need to worry how fast your network-attached storage can feed your GPUs.</p><p>Right?</p><p>Wrong!</p><p>Data distribution over many flash drives enables higher aggregate performance than internal storage. And the time taken to transfer that data over a high speed network is a tiny fraction of the time it takes to perform a read or write operation on a flash drive.</p><p>A modern network-attached storage solution removes the need to copy data into the compute systems. Your GPUs and data scientists are no longer waiting for data.</p><p>This has a positive ‘return-on-investment’ impact. For example, an investment in 10 GPU nodes and 3 data scientists could cost £3M over 3 years. If they’re waiting on data for 30% of the time, then that’s almost £1M wasted.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MyjP541dCpkShGNt6X0ayA.png" /></figure><h3>3. Focusing on bandwidth performance only</h3><p>This is a very common mistake. The performance of file and object storage solutions has often been measured by bandwidth. But bandwidth is only one performance aspect. There are several other aspects including IOPS (I/O Operations/Second), metadata and latency performance.</p><p>That’s why the most popular storage benchmark suites measure more than bandwidth. The IO500, for example, is a league table for the world’s fastest file systems. It has a benchmark suite that measures read and write bandwidth for small and large I/O sizes. It also measures IOPS performance and metadata performance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/550/1*jSBsEPiCtyCHMQ3_38YXMw.png" /><figcaption>IO500 League Table: SuperComputing, November 2019</figcaption></figure><p>If your application handles only large files and sequential I/O operations then measuring bandwidth on its own is fine. But, modern analytics, ML/DL and scientific computing data pipelines are very diverse.</p><p>Here’s an example of the I/O demands in a typical ML/DL data pipeline.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8I-MpTuCrGidGTAApBU5Iw.png" /><figcaption>Diverse I/O Workload in AI Data Pipeline</figcaption></figure><p>Don’t focus only on bandwidth. Understand the I/O performance demands of your data pipeline at each stage. Test storage systems with your most demanding workloads.</p><h3>4. Bottlenecking Single Client Performance with Legacy Protocols</h3><p>NFS and SMB are the most common traditional file storage protocols. NFS was invented in 1984 and SMB in 1983, when the fastest local area network speeds were 4mb/sec. The performance of these protocols was enough at that time and remained that way for decades.</p><p>Both NFS and SMB are performance limited. This wasn’t an issue up to and including 10Gb local area networking. They delivered around that 10Gb/sec performance. But, 100Gb and 200Gb networking is common with today’s GPU compute systems. NFS and SMB have thus become a bottleneck. If you want to saturate your network links with data to keep your GPUs busy, you really need a modern parallel storage protocol.</p><p>NVIDIA also recognises that accelerating data into the GPU nodes is critical. It has released performance enhancements like GPUDirect Storage to improve performance. You can read more about it in my previous blog: ‘ <a href="https://www.derekburke.co.uk/blog/how-gpudirect-storage-accelerates-big-data-analytics">How GPUDirect Storage Accelerates Big Data Analytics’.</a></p><p>It’s important to check what protocols your storage system supports. And what performance it delivers via those protocols to your GPU clients. Check both bandwidth and IOPS performance. If it can’t saturate the network links to your GPU nodes, then it will be the bottleneck for performance. And if you’re investing for 5 or 10 years, that’s a long-time to live with that mistake.</p><p>— — —</p><h3><strong>Summary</strong></h3><p>Avoid these common mistakes, unleash your data and keep those hungry GPUs busy:</p><p>1. Don’t forget about ‘performance per TB’</p><p>2. Don’t waste time copying data</p><p>3. Don’t just focus on bandwidth (remember IOPS &amp; metadata too)</p><p>4. Don’t create client bottlenecks with legacy protocols</p><p>GPU computing and (data) scientists are a huge investment. You are making that investment because it is critical to the success of your organisation. I hope that this information helps you to maximise that success.</p><p>— -</p><p>Thanks for reading my blog. It is my own work and does not represent the views of my employer. I work for a data storage company called WekaIO. Our product, WekaFS is the world’s fastest file system and it’s enabling companies to speed up their data pipelines in scientific and technical computing. You can find out more <a href="https://www.weka.io/">here</a>.</p><p><em>Originally published at </em><a href="http://www.derekburke.co.uk/blog/the-4-worst-mistakes-that-slow-your-gpu-computing-and-how-to-avoid-them"><em>https://www.derekburke.co.uk</em></a><em> on May 1, 2021.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fc9fba21e4d2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Al Trading Algos Learn More to Earn More]]></title>
            <link>https://derekbburke.medium.com/al-trading-algos-learn-more-to-earn-more-f245e1a0b6c6?source=rss-9c1626f353e4------2</link>
            <guid isPermaLink="false">https://medium.com/p/f245e1a0b6c6</guid>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[trading]]></category>
            <dc:creator><![CDATA[Derek Burke]]></dc:creator>
            <pubDate>Thu, 04 Feb 2021 00:00:00 GMT</pubDate>
            <atom:updated>2021-02-04T22:12:49.223Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lpF4BDoOj5_m7RXx" /></figure><p>High Frequency Trading has leveraged arbitrage (taking advantage of the differing prices of an asset to make profit) for years. But, accelerating trade execution is now pushing the boundaries of physics and traders are turning to AI to gain a competitive edge.</p><p>AI is being used to both augment discretional trading and drive systematic trading. For example:</p><p>- Natural language processing for sentiment analysis</p><p>- Reinforcement learning for optimal trade execution</p><p>- Recurrent neural networks for market modelling</p><h3>AI Trading in Action</h3><p>A 2019 <a href="https://www.greenwich.com/market-structure-technology/future-trading-technology-2024">survey by Greenwich Associates</a> stated that:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IbeN1pNxY0BxTzfTArAr9w.png" /></figure><p>There are plenty of examples of AI trading in action.</p><h4>Royal Bank of Canada</h4><p><a href="https://www.rbccm.com/en/expertise/electronic-trading/ai-trading.pageutm_source=rbc&amp;utm_medium=va&amp;utm_campaign=aiden">Aiden, at the Royal Bank of Canada</a>, is an AI-powered electronic trading platform. It uses deep reinforcement learning to improve trading results. Its neural networks can discover new trading relationships through data, anticipate events and take preemptive action. But, the real game-changer is Aiden’s ability to adjust to market conditions. Where previous algorithms had to be fine-tuned offline, Aiden automatically adapts to real-time market changes online.</p><h4>J.P. Morgan</h4><p><a href="https://www.jpmorgan.com/solutions/cib/markets/machine-learning-fx">The Deep Neural Network for Algo Execution (DNA) at J.P. Morgan</a> is leveraging machine learning for FX trading. DNA packages specific existing algorithms into one streamlined execution strategy. It then uses reinforcement learning to measure performance. DNA is presently augmenting existing strategies but, J.P. Morgan’s aim is to create one algorithm that improves execution and adapts to market changes.</p><h4>Kavout</h4><p><a href="https://www.kavout.com/ai-stock-picks/">Kavout</a> leverages pattern recognition technology, price forecasting and equity rating to provide its users with daily stock recommendations. Its AI not only delivers the probability of price rises but also suggests the optimal holding period and the forecasted rate of return.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*u5MeOHstKLqBGrhUmfVsYA.png" /></figure><h4>Epoque Plus</h4><p>AI trading company, <a href="http://www.epoque-plus.ch/">Epoque Plus</a>, provides a fully automated trading service. It specialises in contract for difference (CFDs) on forex, indices, energy &amp; metals, plus SPOT trading on forex on pure currency pairs. In the past year, Epoque Plus’ AI trading service has returned an average monthly performance of 1.55% in a highly volatile market. Risk intelligence is key and Epoque Plus analyses huge amounts of data in real-time to control orders, monitor and anticipate risk.</p><h4>Scotiabank</h4><p><a href="https://www.risk.net/awards/7736276/technology-innovation-of-the-year-scotiabank?ncid=so-nvsh-93347-vt09#cid=ix06_so-nvsh_en-us">Scotiabank</a> has tightened up pricing for its derivatives clients. Its risk engine leverages adjoint algorithmic differentiation (AAD) with deep neural networks to calculate faster and more accurate valuation adjustments (XVAs). Scotiabank claims this combination is a first. And it proved its worth in early 2020 as liquidity evaporated in certain markets due to the coronavirus.</p><h3>Data Growth &amp; Machine Learning</h3><p>Digital data continues to enjoy hyperbolic growth. The finance industry is one of the largest consumers of data and it has more to gain from data analytics than most.</p><p>According to a <a href="https://www.tradersmagazine.com/am/global-spend-on-market-data-totals-record-32b-in-2019/">study by Burton-Taylor International Consulting</a>, the global spend on financial market data grew to $32Bn in 2019 — a 5.6% growth on the previous year.</p><p>A further <a href="https://www.refinitiv.com/perspectives/future-of-investing-trading/the-rising-value-of-data-in-financial-markets/">study in 2019 by Greenwich Associates </a>(in partnership with Refinitiv) stated that 85% of banks, investors and capital market service providers plan to increase spending on data management.</p><p>If data is the new oil, then machine learning is the new refinery. Machine learning requires the data to train algorithms and improve analytical models. Overfitting aside, more data correlates to more accurate results.</p><p>Which leads me to a quote from Warren Buffett, one of the world’s most successful investors:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SAQjWLcMG8BM2g7tlLN5Ug.png" /></figure><p>I’m sure he wasn’t referring to machine learning but, it’s wonderfully applicable — and easier than fighting physics!</p><p>— — — — End — — -</p><p>Thanks for reading my blog. It is my own work and does not represent the views of my employer. But, if you’re interested, I work for a data storage company called WekaIO. Our product, WekaFS, is the world’s fastest file system and it’s enabling financial companies to speed up their data pipelines for trading, backtesting, high performance analytics and machine learning. You can find out more <a href="https://www.weka.io/solutions/financial-services/">here</a>.</p><p>If you like my writing then you can find more of my blogs <a href="https://www.derekburke.co.uk/">here</a>.</p><p><em>Originally published at </em><a href="https://www.derekburke.co.uk/blog/uyo92vsy5ih1utvwrqfd67a383h9do"><em>https://www.derekburke.co.uk</em></a><em> on February 4, 2021.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f245e1a0b6c6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Get 4x Better Storage Price/Performance for Financial Analytics]]></title>
            <link>https://derekbburke.medium.com/how-to-get-4x-better-storage-price-performance-for-financial-analytics-3aef88eb2b22?source=rss-9c1626f353e4------2</link>
            <guid isPermaLink="false">https://medium.com/p/3aef88eb2b22</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[financial-services]]></category>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[tech]]></category>
            <dc:creator><![CDATA[Derek Burke]]></dc:creator>
            <pubDate>Sat, 05 Sep 2020 00:00:00 GMT</pubDate>
            <atom:updated>2020-09-27T17:48:57.835Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Jrc3fVvOu4lUSsuE" /></figure><p>I’ve worked in high performance computing and data storage for 20 years. With a focus (during the past 5 years) on high performance data analytics in finance. And it’s led me to the conclusion that you can improve your storage price/performance by 4x. And you can do it today!</p><p>There are modern storage solutions with better price/performance than your current solution. With software designed to take advantage of:</p><ul><li>NVMe flash storage</li><li>High speed networks (100Gb or 200Gb)</li><li>Data distribution</li><li>Workload parallelism</li></ul><p>I’d like to draw a specific comparison with your incumbent storage solution but, there are too many solutions out there. So, let’s take a high-level look at the pros and cons of the big bucket storage categories. Then I’ll share some direct performance comparisons.</p><h3>Direct Attached Storage (DAS)</h3><h4>The pros:</h4><p>- Low latency performance</p><h4>The cons:</h4><p>- Data is not shared across clients so, you need many copies of the data for client workload parallelism</p><p>- Your DAS system is a single point of failure. You need more copies of the data it holds</p><h3>Storage Area Networks (SAN)</h3><h4>The pros:</h4><p>- Storage resources are shared across multiple clients</p><p>- No single point of failure</p><h4>The cons:</h4><p>- Very expensive</p><p>- A SAN is block-based storage. Modern financial analytics applications (e.g. InfluxDB, MongoDB, KX kdb+, Spark, PyTorch, TensorFlow, SAS Viya, etc.) require a file system</p><p>- Layering file services on top of a block-based SAN adds latency, complexity and cost</p><h3>Network Attached Storage (NAS)</h3><h4>The pros:</h4><p>- Data sharing across many clients</p><p>- No single point of failure</p><h4>The cons:</h4><p>- Shared file system protocols, NFS and SMB, have higher latency than modern protocols</p><p>- NFS and SMB protocols have lower per-client performance than parallel storage</p><p>- NAS appliances offer low performance density</p><h3>Parallel File Systems (aka Parallel NAS)</h3><h4>The pros:</h4><p>- Data sharing across many clients</p><p>- Parallel protocols (specific to the chosen file system) enable higher per-client performance</p><p>- Good performance for large files with sequential IO</p><h4>The cons:</h4><p>- Poor small file and random IO performance</p><p>- Poor metadata performance</p><h3>Introducing <a href="https://www.weka.io/parallel-file-system/">WekaFS</a></h3><p>The Weka File System (WekaFS) is a modern parallel file system designed to unleash the performance of NVMe storage over high-speed networks. It combines the benefits of DAS, NAS and SAN with unparalleled performance, resiliency and cost.</p><p>Let’s take a look at the pros of WekaFS.</p><h4>Performance:</h4><p>- Block storage like latency</p><p>- Up to 3x faster throughput than DAS</p><p>- Parallel protocol saturates 100Gb or 200Gb networks</p><p>- Huge single client performance</p><p>- 4x aggregate performance density versus all-flash NAS appliances</p><p>- Splits all files into 4k shards providing excellent performance for small or large files</p><p>- Distributed metadata and scalable metadata services</p><h4>Resiliency:</h4><p>- Lose up to 4 storage servers without data loss</p><p>- Fast reconstructions in the event of a failed drive</p><p>- Snapshots with zero performance impact</p><p>- Simplified back-up and disaster recovery</p><h4>Total Cost of Ownership:</h4><p>- Ability to leverage commodity hardware</p><p>- Automatic tiering from flash to spinning disk within a shared namespace</p><p>- Simple back-up without third-party back-up software</p><p>- Simple disaster recovery without data replication</p><p>- Public cloud integration for bursting or hybrid deployments</p><h3>Direct Performance Comparisons</h3><p>Finance is a competitive marketplace. Companies do not want to divulge the technology that gives them a competitive edge. So, <a href="https://www.weka.io/">Weka</a> and its partners, HPE and KX, engaged the <a href="https://www.stacresearch.com/">STAC Benchmark Council (STAC)</a>. STAC is a group of leading financial firms and technology vendors. It produces benchmark suites that represent finance application workloads. Weka and its partners chose to run the <a href="https://www.stacresearch.com/m3">STAC M3 Tick Data Analytics benchmark suite</a>.</p><p>STAC M3 has two data sets, Antuco and Kanaga. Antuco is a limited data set that doesn’t stress the storage subsystem. Kanaga is a scaling data set that does.</p><p>Performance is measured in latency for each of the 24 Kanaga benchmarks. The solution is tested with increasing client concurrency and data sets. Of the 24 Kanaga tests, <a href="https://www.weka.io/press-releases/weka-delivers-record-breaking-results-on-17-stac-m3-tick-analytics-benchmarks-using-hpe-apollo-systems/">WekaFS set new world records </a>in 12.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YaMGIIisKRq1dabyeE1i0A.png" /><figcaption>STAC M3 Kanaga Results Comparisons (taken from published results at <a href="https://www.stacresearch.com/">STACresearch.com</a>). Results show latency (lower score is better/faster). World records (at the time of writing) are highlighted in green.</figcaption></figure><p>A comparison of WekaFS results versus competitive solutions (also using KX) draws the following conclusions:</p><p>- WekaFS excels where client concurrency is higher</p><p>- Taking the average latency across all 24 benchmarks WekaFS is:</p><ul><li>3.9x faster than all-flash NAS</li><li>4.5x faster than DAS (Optane SSD Server)</li><li>1.6x faster than SAN (NVMe-oF)</li></ul><h3>Conclusions</h3><p>WekaFS combines the benefits of DAS, NAS and SAN with epic performance for financial analytics.</p><p>WekaFS uses commodity hardware to provide 4x the performance density of all-flash NAS. That’s right! For the same amount of flash you can expect 4x performance.</p><p>— -</p><p>If you’d like to know more then you can visit Weka <a href="https://www.weka.io/">here</a>.</p><p>Sign-up <a href="https://www.derekburke.co.uk/sign-up">here</a> if you’re interested in receiving my blogs directly to your inbox.</p><p><em>Originally published at </em><a href="https://www.derekburke.co.uk/blog/how-to-get-4x-better-storage-priceperformance-for-financial-analytics"><em>https://www.derekburke.co.uk</em></a><em> on September 5, 2020.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3aef88eb2b22" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How GPUDirect Storage Accelerates Big Data Analytics]]></title>
            <link>https://derekbburke.medium.com/how-gpudirect-storage-accelerates-big-data-analytics-1c76bbc72aec?source=rss-9c1626f353e4------2</link>
            <guid isPermaLink="false">https://medium.com/p/1c76bbc72aec</guid>
            <category><![CDATA[gpu-computing]]></category>
            <category><![CDATA[big-data-analytics]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Derek Burke]]></dc:creator>
            <pubDate>Thu, 20 Aug 2020 05:22:09 GMT</pubDate>
            <atom:updated>2020-08-27T08:39:46.808Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*bEQP9xKFkmVHrWht.jpg" /></figure><h3>Performance of GPU Accelerated AI</h3><p>GPU-accelerated computing has traditionally been associated with high performance computing (HPC) and more-recently with <a href="https://www.weka.io/solutions/ai-analytics/">AI/machine learning</a>. However, <a href="https://rapids.ai/">RAPIDS</a> suite of software libraries now enables end-to-end data science and analytics pipelines entirely on GPUs.</p><p>The performance impact of GPU-accelerated computing &amp; RAPIDS was recently illustrated by <a href="https://blogs.nvidia.com/blog/2020/06/22/big-data-analytics-tpcx-bb/">NVIDIA’s unofficial TPCx-BB result</a>, where it beat the previous record by 19.5x! TPCx-BB is a big data benchmark for enterprises representing real-world ETL (extract, transform, load) and machine learning workflows. The benchmark’s 30 queries include big data analytics use cases like inventory management, price analysis, sales analysis, recommendation systems, customer segmentation and sentiment analysis.</p><p>GPU accelerated compute nodes have more than 10x the compute density of CPU-based systems. It is therefore a huge challenge for traditional data storage systems to keep those GPUs busy with data at high performance and low latency. To combat this challenge NVIDIA has a new feature, GPUDirect Storage, that bypasses the CPU &amp; memory enabling GPUs to communicate directly with internal or shared data storage systems.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*A0S9J9UPeg_fu2yr.png" /></figure><p>NVIDIA is an investor in WekaIO (Weka), manufacturer of the Weka File System, <a href="https://www.weka.io/products/">WekaFS</a>. Today, WekaFS is one of only a handful of storage technologies that support GPUDirect storage. WekaFS also has unique capabilities in its ability to handle small files, random IO and metadata-intensive workloads as simply as it handles large files and sequential IO workloads.</p><h3>GPUDirect Storage Results</h3><p>On its recent GPUDirect Storage webinar, NVIDIA chose to illustrate further performance improvements on the TPCx-BB benchmark (separate to the previously referenced results) through the use of GPUDirect Storage and WekaFS. The image below is taken from that webinar and highlights that with WekaFS, GPUDirect Storage provides a further 5x speedup on TPCx-BB results that are the most IO bound.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*63i5MPmwXT6dXjxD.png" /><figcaption>NVIDIA: Accelerating the Data Path to the GPU Webinar</figcaption></figure><p>These results show that the data storage solution should be considered carefully when transitioning to GPU-accelerated computing. Big data analytics can hit the turbo button with GPU acceleration. But, just like a car’s turbo brings more air into the compression chamber, the right data storage solution can feed more data into the GPUs. WekaFS is GPUDirect Storage ready, container-enabled and built for modern analytics with scale-out performance and scale-out capacity on industry-standard hardware.</p><p>WekaIO is seeking beta customers for NVIDIA GPUDirect Storage. If you would like to be considered you can register your interest <a href="https://www.weka.io/get-started">here</a>.</p><p><em>Originally published at </em><a href="https://www.weka.io/blog/how-gpudirect-storage-accelerates-big-data-analytics/"><em>https://www.weka.io</em></a><em> on August 20, 2020.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1c76bbc72aec" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>