Subject: Re: horrible raidframe performance on 2.0-RC5
To: Egervary Gergely <firstname.lastname@example.org>
From: Greg Oster <email@example.com>
Date: 12/03/2004 18:09:26
Egervary Gergely writes:
> I've used raidframe for years now up to 1.6.2. It worked for me
> quite well -
> - about 60-80% performance boost for a simple two-disk stripe,
> - about 80% performance of the single disk for a simple mirror.
> (both are raw, sequential reads/writes)
> now, I tried 2.0-RC5. unfortunately something is very wrong there :(
> raw sequential access looks fair, but random-access is horrible.
> newfs'ing a 65 GB partition takes about 5 minutes (bs=8192,
> fragment=1024) it's about 35-45 seconds on the same hardware w/1.6.2.
Hmm. Same disks too? If you run the newfs on a non-RAID partition,
does it take that long?
> the fs then mounts okay, but copying files onto it takes forever,
> and while it's copying, the box locks up, sometimes even pinging the
> machine, or get into the CRTL-ALT-ESC debugger is impossible.
How much RAM in these boxes? This is sounding like kernel RAM
contention. I'm not sure why copying files would take forever
though... (that would sound more like a disk timeout..)
If you run 'systat iostat', or 'iostat raid0 1', what sort of
throughtput/behaviour do you see? Are there long pauses?
(s/raid0/yourraid/ , of cousre)
> raidframe or other sw raid solution is mission-critical for me, any help
> is greatly appretiated.
> reproduced this on several machines, including HP/Compaq Proliant
> ML370/onboard SCSI/72GB disks and simple desktops. dmesg outputs
> are available, if needed.
Send them to me, please. In particular, we need to figure out
what else has changed in NetBSD since 1.6.2 that might be causing
this. (e.g. other drivers that have changed). I've been running
2.0_BETA w/ RAIDframe in various settings since 2.0_BETA came out,
and have not encountered these sorts of issues.
> on the other side, i386 SMP support is very impressive! guys, keep
> up the good work! but for now, please fix raidframe before 2.0
> hits the markets :)
I suspect 2.0 will ship before I can "fix" RAIDFrame... (especially
since I don't have any way of repeating the problem yet, and it's
probably too late to get anything into 2.0 at this point.)
> simple mirror config:
> START array
> 1 2 0
> START disks
> START layout
> 128 1 1 1
> START queue
> fifo 100
Config looks quite normal.