From: | eric soroos <eric-psql(at)soroos(dot)net> |
---|---|
To: | Curt Sampson <cjs(at)cynic(dot)net> |
Cc: | pgsql-performance(at)postgresql(dot)org |
Subject: | Re: how to configure my new server |
Date: | 2003-02-07 03:07:18 |
Message-ID: | 138182794.1167564058@[4.42.179.151] |
Views: | Raw Message | Whole Thread | Download mbox | Resend email |
Thread: | |
Lists: | pgsql-performance |
On Fri, 7 Feb 2003 10:17:43 +0900 (JST) in message <Pine(dot)NEB(dot)4(dot)51(dot)0302071013180(dot)7356(at)angelic(dot)cynic(dot)net>, Curt Sampson <cjs(at)cynic(dot)net> wrote:
> On Thu, 6 Feb 2003, eric soroos wrote:
>
> > running on a rocketraid card (kernel thinks it's scsi) is faster than
> > onboard controllers
>
> How many transactions per second can you get on a single RAID or stripe
> system using this card?
My current test setup is using a rocketraid 404 card, + 2 WD Caviar 80G SE (8 meg cache), and a 2 yr old 7200 rpm ide ibm on the mb controller channel as the os/log drive. The rocketraid is a 4 channel card, I'm going to fill the other two channels with mirror drives. I'm _not_ using hw mirroring, since I want the flexibility of sw mirroring for now. (I don't think their driver supports breaking and reestablishing the mirror with live drive usage)
This is on a single p3-733, 640 mb ram. Processor is apparently never redlined normally running at 50-75% as reported by vmstat.
> I found that the write performance for large writes on an Escalade 7850
> was great, but I couldn't coax more than about 120 writes per second
> out of the thing in any configuration (even writing to separate disks
> in JBOD mode), which made it very disappointing for database use. (The
> individual disks on the controller, modern IBM 60 GB IDEs, could do
> about 90 writes per second.)
Using a data/wal split, I see peaks around 135 t/s (500 transactions, 10 concurrent clients). Sustained (25000 transactions) that goes down to the 80 range due to the moving of data from the WAL to the data directories. these numbers are all with the kernel's data caches full of about 1/2 gig of data, I'm seeing 300 blks/s read and 3000 write
Peaks for striped are @ 80, peaks for single are ~ 100, peaks for mirror are around 100. I'm curious if hw mirroring would help, as I am about a 4 disk raid 5. But I'm not likely to have the proper drives for that in time to do the testing, and I like the ablity to break the mirror for a backup. For comparison, that same system was doing 20-50 without the extra 512 stick of ram and on the internal single drive.
eric
From | Date | Subject | |
---|---|---|---|
Next Message | Andreas Pflug | 2003-02-07 03:23:34 | Re: how to configure my new server |
Previous Message | Curt Sampson | 2003-02-07 01:17:43 | Re: how to configure my new server |