On 29/05/2012 11:10 PM, Jim Klimov wrote:
2012-05-29 16:35, Nathan Kroenert wrote:
Hi John,
Actually, last time I tried the whole AF (4k) thing, it's performance
was worse than woeful.
But admittedly, that was a little while ago.
The drives were the seagate green barracuda IIRC, and performance for
just about everything was 20MB/s per spindle or worse, when it should
have been closer to 100MB/s when streaming. Things were worse still when
doing random...
On one hand, it is possible that being green, the drives aren't very
capable of fast IO - they had different design goals and tradeoffs.
Indeed! I just wasn't expecting it to be so profound.
But actually I was going to ask if you paid attention to partitioning?
At what offsets did your ZFS pool data start? Was that offset divisible
by 4KB (i.e. 256 512byte sectors as is default now vs 34 sectors of
the older default)?
It was. Actually I tried it in a variety of ways, including auto EFI
partition (zpool create with the whole disk), using SMI label, and
trying a variety of tricks with offsets. Again, it was a while ago -
before the time of the SD RMW fix...
If the drive had 4kb native sectors but the logical FS blocks were
not aligned with that, then every write IO would involve RMW of
many sectors (perhaps disk's caching might alleviate this for
streaming writes though).
Yep - that's what it *felt* like, and I didn't seem to be able to change
that at the time.
Also note that ZFS IO often is random even for reads, since you
have to read metadata and file data often from different dispersed
locations. Again, OS caching helps statistically, when you have
much RAM dedicated to caching. Hmmm... did you use dedup in those
tests?- that is another source of performance degradation on smaller
machines (under tens of GBs of RAM).
At the time, I had 1TB of data, and 1TB of space... I'd expect that most
of the data would have been written 'closeish' to sequential on disk,
though I'll confess I only spent a short time looking at the 'physical'
read/write locations being send down through the stack. (where the drive
writes them - well.. That's different. ;)
I have been contacted off list by a few folks that have indicated
success with current drives and current Solaris bits. I'm thinking that
it might be time to take another run at it.
I'll let the list know the results. ;)
Cheers
Nathan.
_______________________________________________
zfs-discuss mailing list
zfs-discuss@opensolaris.org
http://mail.opensolaris.org/mailman/listinfo/zfs-discuss
