[DRBD-user] 0.6 vs. 0.7 3ware performance? [Was: Expected syncspeed for a 425gb volume?]

Jeff Buck jeffb at umci.com
Tue Apr 12 19:57:53 CEST 2005

Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.


I'm a little late on this one, but I'll chime in anyway.. 

We too have had serious issues with performance + 3ware (85xx) + drbd
(7). We've tried tweaking the crap out of it also, and had no luck. 

After talking to 3ware, they seem to think that part of the problem had
to do with the drives we're using.. We've got 8 250 gig western digital
drives. The problem is the particular model (according to them). Some of
the western digital drives are "workstation class" drives, and when they
see info coming in in 64K chunks, they assume it's some sort of nickle
and dime IO from a workstation, so after a 64k chunk they seek to a
random area of the disk (ala IBM Deathstar drive fix) to even out the
time spent over different areas on the disk. Next 64k chunk they seek
back, write, and seek random again. The problem being that the 3ware
raid card is splitting the writes into 64k chunks on the different
disks, causing them all to do a little dance after every 64k write. We
have a bunch of the "JD" drives.. I guess the ones that don't have this
behavior are the "SD" drives (unavailable when we started this
project). 

Perhaps drbd 7 writes in different size chunks aggravating a crappy
3ware -> drive problem?

A while back /. linked to  a review of a bunch of raid cards, and it was
somewhat enlightening. The 3ware really didn't do very well. Some of the
problems we noticed when reviewing the article:

Raid 5 performance actually drops as you add more drives (a 4 drive
array is much faster than an 8 drive array, but gets it's butt kicked by
a single drive).

On the performance front, the 3wares _really_ had their ass handed back
to them by other comparably priced (or cheaper) cards.

We decided to just give up, and we bought a couple of Aereca cards.
Unfortunately we don't have any performance tests yet, we just got the
cards.

Is there anyone out there getting decent performance from 3ware cards +
drbd 7?


On Thu, 2005-04-07 at 00:43 +0200, Felix Ide wrote:
> With our two 3ware 7500 controllers write performance with 0.7 was so
> bad that the system was nearly unusuable (with ~100 active users on
> samba shares). With 0.6 (running since 1 1/2 years IIRC) performance is
> quite good (don't have benchmarks at hand right now, the machines with
> documentation are on a customers site, could check tomorrow).
> If I can help debugging, I could do some tests on friday afternoon.
> 
> Felix 
> 
> Am Donnerstag, den 07.04.2005, 09:45 +1200 schrieb James Doherty:
> > On Wed, 2005-04-06 at 15:49 -0500, Nate Carlson wrote:
> > > On Wed, 6 Apr 2005, Todd Denniston wrote:
> > > > Ouch, that is one heck of a diffrence.  However on thing 0.7 has over
> > > > 0.6 is that even on a cold restart, once the devices have fully synced
> > > > the first time, is that unless you force it to, 0.7 only has to sync
> > > > what has changed.
> > > 
> > > Yeah, but the slow write speed for normal operations won't be acceptable.  
> > > :(
> > 
> > It's interesting that you should point that out. We have a RAID5 array
> > with a 3ware 7500-4 in one machine and 3ware 3W-6400 in the other. Both
> > machines have 3com gigabit nic's. We've only ever used DRBD 0.7.5 on
> > this setup and sync performance isn't too shabby (20,000 K/sec, although
> > I think I may have seen 30,000) but write performance seems to be rather
> > poor (no benchmarks on this).
> > 
> > The machines aren't the fastest machines around, but they're fairly
> > decent Athlon XP's with lots of ram. Perhaps we are at the limits of our
> > hardware, but I expected a little better performance while writing to
> > the drbd device. Perhaps it is a 3ware<->DRBD issue?
> > 
> > Cheers
> 
> _______________________________________________
> drbd-user mailing list
> drbd-user at lists.linbit.com
> http://lists.linbit.com/mailman/listinfo/drbd-user



More information about the drbd-user mailing list