[OpenAFS] improving cache partition performance
Mon, 29 Aug 2011 18:42:37 -0400
Both of my servers have 16GB of RAM, and yes, I'm serving PHP content.
I'm not sure if I can add more spindles in these boxes with my RAID
controller. (1U Dell PowerEdge R210 with H200 RAID card.)
I disabled APC cache because of a segfault every ~5000 hits, which
happened to hit during an important demo by my boss.
I was asking about commit time to see if there was a way to combine disk
transactions by using more RAM, and to see how dangerous that might be.
On 08/28/2011 11:00 PM, Chris Hoy Poy wrote:
> Hi Jason,
> RAM - cache more.
> Otherwise - add more spindles.
> Improve the local AFS cache for each webserver - is that local or disk ?
> I wouldn't mess around with the EXT3 settings, due to the way OpenAFS works. In my experience, messing with these settings - you might get a touch more output, but the time associated with it, compared with the cost of adding more spindles or adding more RAM, is typically not worth it. You aren't likely to get more then a few ms or a few more IOPS - either way. Best to push it out of RAM altogether.
> Other things to check - are your operations READ bottlenecked or WRITE bottlenecked? Whats your proportion of read to write operations? If you are heavy on reads - definitely get more RAM into the system. If you are write bottlenecked - you need more spindles (though if your size is ~50-100GB, have a look at the SSD options - you can get some excellent raided SSDs fairly cheap these days).
> If the webserver is serving PHP content, look at APC cache to get (again) get the reads off the disk.
> ----- Original Message -----
> From: "Jason Edgecombe"<email@example.com>
> To: firstname.lastname@example.org
> Sent: Monday, 29 August, 2011 5:00:23 AM
> Subject: [OpenAFS] improving cache partition performance
> hi everyone,
> I'm looking at some performance problems on our web servers, and I think
> that hard drive performance is a bottleneck. All filesystems are on a
> RAID0 disk array with two 7200RPM disks (no battery-backed cache). The
> cache is on a dedicated ext3 partition. At some points, my await times
> on the disk are up to 1000ms. How can I improve this? Would setting
> commit time larger than 5 seconds be OK? Would setting the ext3 journal
> type to "data=writeback" be OK?
> platform: RHEL5.5 x86_64
> _______________________________________________ OpenAFS-info mailing
> list OpenAFSemail@example.com