On Dec 18, 2008, at 12:24 PM, "John" jses27@gmail.com wrote:
-----Original Message----- From: centos-bounces@centos.org [mailto:centos-bounces@centos.org] On Behalf Of Matt Sent: Thursday, December 18, 2008 10:37 AM To: CentOS mailing list Subject: Re: [CentOS] Adding RAM
A bit of bottle neck.
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s
rkB/s wkB/s
avgrq-sz avgqu-sz await svctm %util sda 0.38 176.63 70.32 78.26 813.46 2044.82
406.73 1022.41
19.24 0.40 19.17 4.04 60.07 sda1 0.00 0.00 0.00 0.00 0.00 0.00
0.00 0.00
5.28 0.00 23.61 19.33 0.00 sda2 0.38 176.63 70.32 78.26 813.45 2044.82
406.73 1022.41
19.24 0.40 19.17 4.04 60.07 dm-0 0.00 0.00 70.71 255.60 813.45 2044.82 406.73 1022.41 8.76 2.90 8.87 1.84 60.10 dm-1 0.00 0.00 0.00 0.00 0.00 0.00
0.00 0.00
8.00 0.00 64.20 11.38 0.00
Try setting the scheduler to 'deadline' and see if the queue sizes shrink.
No raid1? Besides adding redundancy, it can help with read performance. I would probably put the mail on a raid 10 though if I had 4 disks to do so.
-Ross
Like this:
kernel /vmlinuz-2.6.9-78.0.8.ELsmp ro
root=/dev/VolGroup00/LogVol00 elevator=deadline
And this change will be for System Wide.
Doing this among other things helped quite a lot. My iostat -x %util dropped from ~ 60% to ~ 22% now. Of course at same time I updated from dual-core 2Ghz CPU to quad-core 2.4Ghz. Also swapped motherboard from one with an "Intel ICH9R" Southbridge to a "Intel ICH7R" since I heard CentOS 4.x did not have drivers in kernel for ICH9R.
That's why I asked what kind of Controler the board had on it in a previous post to you and stated ram was not the suspect problem. IMO if you keeped the dual core proc and just switched to ICH7 Board you would have saved money. Your utilization rate would probly stayed the same or no higher than %30. Just to keep things in balance you will probly want to try the cfq schedular with a high user load so every thing gets it fair share in time_wait. Some people will contradict that it's about making the users happy. When access time for one user takes longer than another then the complaints start coming in. I would like to know the Proc Utilization per core or are you running it in Single Core?
CFQ accounting is only good for interactive processes, so unless of is a terminal server mucking with CFQ will have little affect.
Deadline and anticipatory schedulers work best for non-interactive servers such as file, mail, database. Noop works for special apps that like to do their own io scheduling, database systems with their own io schedulers mostly.
So it isn't one scheduler to rule them all, but one for each situation.
-Ross