<div dir="ltr">I know that QDR IB gives the best bang for buck currently and that's what we have now. However due to various reasons we are looking at alternatives hence the question. Thank you very much for your information, Ben.</div><div class="gmail_extra"><br><div class="gmail_quote">On 19 June 2015 at 16:24, Ben Evans <span dir="ltr"><<a href="mailto:bevans@cray.com" target="_blank">bevans@cray.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<div lang="EN-US" link="blue" vlink="purple">
<div>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d">It’s faster in that you eliminate all the TCP overhead and latency. (something on the order of 20% improvement in speed, IIRC, it’s been several years)<u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d"><u></u> <u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d">Balancing your network performance with what your disks can provide is a whole other level of system design and implementation. You can stack enough disks
or SSDs behind a server so that the network is your bottleneck, you can stack up enough network to few enough disks so that the drives are your bottleneck. You can stack up enough of both so that the PCIE bus is your bottleneck.<u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d"><u></u> <u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d">Take the time and compare costs/performance to Infiniband, since most systems have a dedicated client/server network, you might as well go as fast as you can.<u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d"><u></u> <u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d">-Ben Evans<u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d"><u></u> <u></u></span></p>
<p class="MsoNormal"><b><span style="font-size:10.0pt;font-family:"Tahoma","sans-serif"">From:</span></b><span style="font-size:10.0pt;font-family:"Tahoma","sans-serif""> <a href="mailto:igko50@gmail.com" target="_blank">igko50@gmail.com</a> [mailto:<a href="mailto:igko50@gmail.com" target="_blank">igko50@gmail.com</a>]
<b>On Behalf Of </b>INKozin<br>
<b>Sent:</b> Friday, June 19, 2015 11:10 AM<br>
<b>To:</b> Ben Evans<br>
<b>Cc:</b> <a href="mailto:lustre-discuss@lists.lustre.org" target="_blank">lustre-discuss@lists.lustre.org</a><br>
<b>Subject:</b> Re: [lustre-discuss] Lustre over 10 Gb Ethernet with and without RDMA<u></u><u></u></span></p><div><div class="h5">
<p class="MsoNormal"><u></u> <u></u></p>
<div>
<p class="MsoNormal">Ben, is it possible to quantify "faster"?<u></u><u></u></p>
<div>
<p class="MsoNormal">Understandably, for a single client on an empty cluster it may feel "faster" but on a busy cluster with many reads and writes in flight I'd have thought the limiting factor is the back end's throughput rather than the network, no? As long
as the bandwidth to a client is somewhat higher than the average i/o bandwidth (back end's throughput divided by the number of clients) the client should be content.<u></u><u></u></p>
</div>
</div>
<div>
<p class="MsoNormal"><u></u> <u></u></p>
<div>
<p class="MsoNormal">On 19 June 2015 at 14:46, Ben Evans <<a href="mailto:bevans@cray.com" target="_blank">bevans@cray.com</a>> wrote:<u></u><u></u></p>
<div>
<div>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d">It is faster, but I don’t know what price/performance tradeoff is, as I only used it as an engineer.</span><u></u><u></u></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d"> </span><u></u><u></u></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d">As an alternative, take a look at RoCE, it does much the same thing but uses normal (?) hardware.
It’s still pretty new, though, so you might have some speedbumps.</span><u></u><u></u></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d"> </span><u></u><u></u></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d">-Ben Evans</span><u></u><u></u></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d"> </span><u></u><u></u></p>
<p class="MsoNormal"><b><span style="font-size:10.0pt;font-family:"Tahoma","sans-serif"">From:</span></b><span style="font-size:10.0pt;font-family:"Tahoma","sans-serif""> lustre-discuss [mailto:<a href="mailto:lustre-discuss-bounces@lists.lustre.org" target="_blank">lustre-discuss-bounces@lists.lustre.org</a>]
<b>On Behalf Of </b>INKozin<br>
<b>Sent:</b> Friday, June 19, 2015 5:43 AM<br>
<b>To:</b> <a href="mailto:lustre-discuss@lists.lustre.org" target="_blank">lustre-discuss@lists.lustre.org</a><br>
<b>Subject:</b> [lustre-discuss] Lustre over 10 Gb Ethernet with and without RDMA</span><u></u><u></u></p>
<div>
<div>
<p class="MsoNormal"> <u></u><u></u></p>
<div>
<div>
<p class="MsoNormal">My question is about performance advantages of Lustre RDMA over 10 Gb Ethernet. When using 10 Gb Ethernet to build Lustre, is it worth paying the premium for iWARP? I understand
that iWARP essentially reduces latency but less sure of its specific implications for storage. Would it improve performance on small files? Any pointers to representative benchmarks will be very appreciated.<u></u><u></u></p>
</div>
<div>
<p class="MsoNormal"> <u></u><u></u></p>
</div>
<div>
<div>
<p class="MsoNormal">Celsio has released a white paper in which they compare Lustre RDMA over 40 Gb Ethernet and FDR IB<u></u><u></u></p>
</div>
<p class="MsoNormal"><a href="http://www.chelsio.com/wp-content/uploads/resources/Lustre-Over-iWARP-vs-IB-FDR.pdf" target="_blank">http://www.chelsio.com/wp-content/uploads/resources/Lustre-Over-iWARP-vs-IB-FDR.pdf</a><u></u><u></u></p>
<div>
<p class="MsoNormal">where they claim comparable performance of both. <u></u><u></u></p>
</div>
</div>
<div>
<p class="MsoNormal">How much worse the throughput on small block sizes would be without iWARP?<u></u><u></u></p>
</div>
<div>
<p class="MsoNormal"> <u></u><u></u></p>
</div>
<div>
<p class="MsoNormal">Thank you<u></u><u></u></p>
</div>
<div>
<p class="MsoNormal">Igor <u></u><u></u></p>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
<p class="MsoNormal"><u></u> <u></u></p>
</div>
</div></div></div>
</div>
</blockquote></div><br></div>