[gpfsug-discuss] Building GPFS filesystem system data pool on shared nothing NVMe drives

Tue Jul 30 14:03:54 BST 2019

OK, so the ECE recovery group is the four NSD servers with the System storage pool disks, and somehow I have to read the docs
and find out how to define pdisks that spread the replication across the four servers, but three disks at a time.  
Three pdisks of 7 drives, three I can't do anything with, or are those for rebuilding space?

Can you provide me details of your six-node non-ECE configuration?  Basically how the NSDs are defined...

The remainder of our new filesystem will have a fast pool of 12 nodes of excelero, and 2Pb of spinning disks, so another possibility
would be to license four more nodes and put the system pool under excelero.  

 -- ddj

> On Jul 30, 2019, at 8:19 AM, Sanchez, Paul <Paul.Sanchez at deshaw.com> wrote:
> 
> Hi David,
>  
> In an ECE configuration, it would be typical to put all of the NVMe disks in all 4 of your servers into a single recovery group.   So in your case, all 24 NVMe drives would be in one recovery group and the 4 servers would be “log group” servers in the recovery group, distributing the I/O load for the NSD/vdisks that are hosted on the RG.  (The minimum disks for a single RG config is 12, and you meet that easily.)
>  
> https://www.ibm.com/support/knowledgecenter/STXKQY_ECE_5.0.3/com.ibm.spectrum.scale.ece.v5r03.doc/b1lece_plan_recommendations.htm <https://www.ibm.com/support/knowledgecenter/STXKQY_ECE_5.0.3/com.ibm.spectrum.scale.ece.v5r03.doc/b1lece_plan_recommendations.htm>
> outlines the recommendations for raidCode protection.  Your configuration (4 nodes) would use vdisks with 4+3P, which gives you a slightly better capacity yield than RAID10 would, but with much better recovery characteristics:
>  
> ·         No single failed node will result in a down system NSD.
> ·         No single drive failure will require a critical priority rebuild, and can be handled in the background without killing performance.
>  
> So from that perspective, ECE is a win here and avoids a problem with the non-ECE, shared-nothing designs: the manual “mmchdisk <fsname> start -a” operation that is needed after any traditional shared-nothing metadata NSD goes offline to bring it back and protect against further failures.
>  
> Despite the operational challenges of the non-ECE design, it can sometimes survive two server failures (if replication factor is 3 and the filesystem descriptor quorum wasn’t lost by the two failures) which a 4 node ECE cluster cannot.  Given that the world is complex and unexpected things can happen, I’d personally recommend redistributing the 24 disks across 6 servers if you can, so that the design could always survive 2 node failures.  I’ve run this design and it’s fairly robust.
>  
> In any event, you should of course test the failure scenarios yourself before going into production to validate them and familiarize yourself with the process.  And a special note on ECE: due to the cooperative nature at the pdisk level, the network between the servers in the RG should be as reliable as possible and any network redundancy should also be tested ahead of time.
>  
> -Paul
>  
> From: gpfsug-discuss-bounces at spectrumscale.org <mailto:gpfsug-discuss-bounces at spectrumscale.org> <gpfsug-discuss-bounces at spectrumscale.org <mailto:gpfsug-discuss-bounces at spectrumscale.org>> On Behalf Of David Johnson
> Sent: Tuesday, July 30, 2019 7:46 AM
> To: gpfsug main discussion list <gpfsug-discuss at spectrumscale.org <mailto:gpfsug-discuss at spectrumscale.org>>
> Subject: Re: [gpfsug-discuss] Building GPFS filesystem system data pool on shared nothing NVMe drives
>  
> This message was sent by an external party.
> 
>  
> Can we confirm the requirement for disks per RG?  I have 4 RG, but only 6 x 3TB NVMe drives per box.
> 
> 
> On Jul 29, 2019, at 1:34 PM, Luis Bolinches <luis.bolinches at fi.ibm.com <mailto:luis.bolinches at fi.ibm.com>> wrote:
>  
> Hi, from phone so sorry for typos.  
>  
> I really think you should look into Spectrum Scale Erasure Code Edition (ECE) for this. 
>  
> Sure you could do a RAID on each node as you mention here but that sounds like a lot of waste to me on storage capacity. Not to forget you get other goodies like end to end checksum and rapid rebuilds with ECE, among others. 
>  
> Four servers is the minimum requirement for ECE (4+3p) and from top of my head 12 disk per RG, you are fine with both requirements. 
> 
> There is a presentation on ECE on the user group web page from London May 2019 were we talk about ECE. 
>  
> And the ibm page of the product https://www.ibm.com/support/knowledgecenter/STXKQY_ECE_5.0.3/com.ibm.spectrum.scale.ece.v5r03.doc/b1lece_intro.htm <https://www.ibm.com/support/knowledgecenter/STXKQY_ECE_5.0.3/com.ibm.spectrum.scale.ece.v5r03.doc/b1lece_intro.htm>
> -- 
> Cheers
> 
> El 29 jul 2019, a las 19:06, David Johnson <david_johnson at brown.edu <mailto:david_johnson at brown.edu>> escribió:
> 
> We are planning a 5.0.x upgrade onto new hardware to make use of the new 5.x GPFS features.
> The goal is to use up to four NSD nodes for metadata, each one with 6 NVMe drives (to be determined
> whether we use Intel VROC for raid 5 or raid 1, or just straight disks).  
> 
> So questions — 
> Has anyone done system pool on shared nothing cluster?  How did you set it up?
> With default metadata replication set at 3, can you make use of four NSD nodes effectively?
> How would one design the location vectors and failure groups so that the system metadata is
> spread evenly across the four servers?
> 
> Thanks,
> — ddj
> Dave Johnson
> _______________________________________________
> gpfsug-discuss mailing list
> gpfsug-discuss at spectrumscale.org <http://spectrumscale.org/>
> http://gpfsug.org/mailman/listinfo/gpfsug-discuss <http://gpfsug.org/mailman/listinfo/gpfsug-discuss>
> 
> Ellei edellä ole toisin mainittu: / Unless stated otherwise above:
> Oy IBM Finland Ab
> PL 265, 00101 Helsinki, Finland
> Business ID, Y-tunnus: 0195876-3 
> Registered in Finland
> 
> _______________________________________________
> gpfsug-discuss mailing list
> gpfsug-discuss at spectrumscale.org <http://spectrumscale.org/>
> http://gpfsug.org/mailman/listinfo/gpfsug-discuss <http://gpfsug.org/mailman/listinfo/gpfsug-discuss>
>  
> _______________________________________________
> gpfsug-discuss mailing list
> gpfsug-discuss at spectrumscale.org <http://spectrumscale.org/>
> http://gpfsug.org/mailman/listinfo/gpfsug-discuss <http://gpfsug.org/mailman/listinfo/gpfsug-discuss>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://gpfsug.org/pipermail/gpfsug-discuss_gpfsug.org/attachments/20190730/e8dc0611/attachment-0002.htm>