2 Jun
2004
2 Jun
'04
2:32 a.m.
Joel White wrote:
Malcolm Baldridge wrote:
Wow, an 8 bad mistake! Actually, I always kinda wondered about that.
This could actually explain a lot of my latencytest results: diskread
latency performance was fine, but diskwrite and diskcopy was crummy.
Ok, so I dug up an ancient 0.5G Conner IDE drive I had lying around
and kludged it in. Clocks in at a paltry 2MB/s transfer rate with
hdparm. I was able to record 8 channels with Ardour with JACK set at
128 periods with no xrun in sight - yeehaa! I was about to plan on
picking up an IDE/ATA drive tomorrow, but now I have something else to
try first.
Thanks, Malcolm.
Joel
Ok, I busted up my RAID and with high hopes fired up jack and ardour on
a new ext3 filesystem running on a single partition. Ardour ran for
about 7 sec and stopped with a 300+ms xrun. This was running 8 inputs,
and even running two inputs, things ground to a halt after about a minute.
Just to be sure, I mounted the old Connor IDE and let thing ardour run.
Got bored by about 3 min and shut it off. Buying a new drive is looking
attractive again...
Thanks for the suggestions. I may still look into reiserfs and the LCK1
patch set, but I don't have high hopes (a 300+ms xrun?). I'm thinking
the aforementioned problems with the Adaptec 2940 is the likely root of
my troubles.
Joel
I'm
using 4 Seagate SX118273LC (18G) drives in a software RAID-1,
mostly
because I wanted to see what putting together a RAID was like (and
at the time I bought them, the drives were relatively cheap).
Bad bad bad bad bad bad bad bad. RAID1 is good for fault tolerance,
and
for read performance... but worse than a single drive's performance for
randomised writes. Software RAID implementations would give you worse
performance on writes under all conditions, I wager.
Try a non-RAID volume in your tests. Keep in
mind that one write is
triggering TWO writes in TWO transactions over the same SCSI card
(and bus,
probably). This isn't going to be good for latency or bus utilisation.
Before you rip out too much hardware (and hair), try it with a "simple"
volume first.
=MB=