<div dir="ltr"><div><div><div><div><div>We never had a budget sufficient to buy a large SSD-only storage so can't say anything about that.<br></div>Our NFS filers tend to crumple when users start genuinely running jobs against them and this tends to affect both the jobs and also any other processes/users who are trying to use the filers.<br></div>Of course it depends on the load type where a single IO process can get really good performance but once you have multiple threads doing IO to different files the NFS servers start to thrash.<br></div>Of course this is also possible with an underspecced Parallel filesystem but so far we have had far less issues with Lustre.<br></div>HTH,<br></div>Eli<br><div><div><div><br></div></div></div></div><div class="gmail_extra"><br><div class="gmail_quote">On Tue, May 1, 2018 at 1:10 AM, Dilger, Andreas <span dir="ltr"><<a href="mailto:andreas.dilger@intel.com" target="_blank">andreas.dilger@intel.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><span class="">On Apr 30, 2018, at 07:11, Thackeray, Neil L <<a href="mailto:neilt@illinois.edu">neilt@illinois.edu</a>> wrote:<br>
> <br>
> Sorry, I left out file size. We don't foresee growing tremendously. The plan is for researchers to upload their data, get the results, and copy it down to a mounted file system. This is going to be used by multiple researchers, and we will be charging for compute time. We really don't want this cluster to be used for storing data outside of the time needed for their computations. We may just start with 100TB of SSD storage.<br>
<br>
</span>One of the major benefits of Lustre is that it can be used directly for large-scale computing. Having users copy data to/from Lustre is fairly inefficient (though surprisingly copying files to/from a direct Lustre mount can be faster than FTP or SCP or other network copy tools).<br>
<br>
You'd be better off to increase the size of your Lustre filesystem, enough that users can store "projects" there for some time while they compute, rather than needing to move the data on/off the filesystem a lot.<br>
<br>
While using an all-SSD filesystem is appealing, you might find better performance with some kind of hybrid storage, like ZFS + L2ARC + Metadata Allocation Class (this feature is in development, target 2018-09, depending on your timeframe). <br>
<br>
You definitely want your MDT(s) to be SSDs, especially if you use the new Data-on-MDT feature to store small files tehre. The OSTs can be HDDs to give you a lot more capacity for the same price.<br>
<br>
Cheers, Andreas<br>
<div><div class="h5"><br>
> -----Original Message-----<br>
> From: lustre-discuss <<a href="mailto:lustre-discuss-bounces@lists.lustre.org">lustre-discuss-bounces@lists.<wbr>lustre.org</a>> On Behalf Of Philippe Weill<br>
> Sent: Saturday, April 28, 2018 1:14 AM<br>
> To: <a href="mailto:lustre-discuss@lists.lustre.org">lustre-discuss@lists.lustre.<wbr>org</a><br>
> Subject: Re: [lustre-discuss] Do I need Lustre?<br>
> <br>
> <br>
> <br>
> Le 27/04/2018 à 19:07, Thackeray, Neil L a écrit :<br>
>> I\u2019m new to the cluster realm, so I\u2019m hoping for some good advice. We <br>
>> are starting up a new cluster, and I\u2019ve noticed that lustre seems to be used widely in datacenters. The thing is I\u2019m not sure the scale of our cluster will need it.<br>
>> <br>
>> We are planning a small cluster, starting with 6 -8 nodes with 2 GPUs <br>
>> per node. They will be used for Deep Learning, MRI data processing, <br>
>> and Matlab among other things. With the size of the cluster we figure <br>
>> that 10Gb networking will be sufficient. We aren\u2019t going to allow persistent storage on the cluster. Users will just upload and download data. I\u2019m mostly concerned about I/O speeds. I don\u2019t know if NFS would be fast enough to handle the data.<br>
>> <br>
>> We are hoping that the cluster will grow over time. We are already talking about buying more nodes next fiscal year.<br>
>> <br>
>> Thanks.<br>
>> <br>
> <br>
> hello<br>
> <br>
> you didn't say anything about filesystem size needed and if you are thinking to grow fast we also run a small cluster ( 20 nodes ) but for climate data modeling results and satellite atmospheric data analysis we are growing at least 300TB per year (2PB now) and it's easier for us to grow with lustre<br>
> <br>
> <br>
> --<br>
> Weill Philippe - Administrateur Systeme et Reseaux<br>
> CNRS/UPMC/IPSL LATMOS (UMR 8190)<br>
> ______________________________<wbr>_________________<br>
> lustre-discuss mailing list<br>
> <a href="mailto:lustre-discuss@lists.lustre.org">lustre-discuss@lists.lustre.<wbr>org</a><br>
> <a href="http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org" rel="noreferrer" target="_blank">http://lists.lustre.org/<wbr>listinfo.cgi/lustre-discuss-<wbr>lustre.org</a><br>
> ______________________________<wbr>_________________<br>
> lustre-discuss mailing list<br>
> <a href="mailto:lustre-discuss@lists.lustre.org">lustre-discuss@lists.lustre.<wbr>org</a><br>
> <a href="http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org" rel="noreferrer" target="_blank">http://lists.lustre.org/<wbr>listinfo.cgi/lustre-discuss-<wbr>lustre.org</a><br>
<br>
</div></div>Cheers, Andreas<br>
--<br>
Andreas Dilger<br>
Lustre Principal Architect<br>
Intel Corporation<br>
<div class="HOEnZb"><div class="h5"><br>
<br>
<br>
<br>
<br>
<br>
<br>
______________________________<wbr>_________________<br>
lustre-discuss mailing list<br>
<a href="mailto:lustre-discuss@lists.lustre.org">lustre-discuss@lists.lustre.<wbr>org</a><br>
<a href="http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org" rel="noreferrer" target="_blank">http://lists.lustre.org/<wbr>listinfo.cgi/lustre-discuss-<wbr>lustre.org</a><br>
</div></div></blockquote></div><br></div>