[gpfsug-discuss] Flash for metadata

[Simon Thompson (Research Computing - IT Services)]

Thanks to all those who replied (including off list).

Someone reminded me that whilst the FlashSystem will support 512 and 4k blocks, it "prefers" 4k. Maybe there is a performance penalty for the smaller writes.

Second, a flash update is always going to be read erase write.

We don't get any choice with FlashSystem 900 on the RAID set - only RAID5 is available.

Given we have two FlashSystems in two data centres, looks like we are going with 4k sectors and GPFS replication.

I assume that creating several smaller LUNs on each FlashSystem in the same failure group is still preferable to one big LUN so we get more IO queues to play with?

Simon

-----Original Message-----
From: Marc A Kaplan [makaplan at us.ibm.com<mailto:makaplan at us.ibm.com>]
Sent: Wednesday, May 18, 2016 12:27 AM GMT Standard Time
To: gpfsug main discussion list
Subject: Re: [gpfsug-discuss] Flash for metadata

I think there are two points here:

A) RAID striping is probably a "loser" for GPFS metadata.
B) RAID mirroring for your metadata may or may not be faster and/or more reliable than GPFS replication. Depending on your requirements and assumptions for fault-tolerance one or the other might be the winner.

But .. if you're serious about performance - measure, measure, measure and pick the winner. (And we'd like to know what you found!)
And if you want the best performance, you'll probably need to (re)measure whenever you acquire new equipment.

GPFS can do and does striping for high performance when reading and writing metadata and data...

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://gpfsug.org/pipermail/gpfsug-discuss_gpfsug.org/attachments/20160519/fdec78f4/attachment-0002.htm>