Re: Watchdog feature in Polygraph-2.8.0

On Wed, 2005/02/09 (MST), <[email protected]> wrote:

> I want to test how file system performance can affect Squid Web Cache
> performance. I decided to test the following file systems: ext3, ext2
> (with options: noatime, async), reiserfs and xfs.

Are you familiar with Duane Wessels' work on Squid disk performance? Do  
you have access to Duane's Squid book[1]? In Appendix D, Duane describes  
many Squid/Polygraph filesystem tests and uses a workload similar to  
yours...

   [1] http://www.squidbook.org/

> Problem is that I don't
> know how to write appropriate workload. I want polyclt robots to increase
> gradually request rate until DUT achieve maximum of acceptable value of
> response time. Then polyclt needs to adjusts local population of robots  
> in order to sustain an acceptable value of response time and finally  
> fill the cache twice.

Overall, your workload makes sense to me. Specific sample sizes, load  
levels, and response time levels may need to be tuned further to achieve  
the desired effect, but you are on the right track.

> However, test like that should shows that file system have influence on  
> Proxy performance. Unfortunately my results don't confirm above
> statement.

There are a few possible explanations, I think. Here are some, in no  
particular order. Note that I am not asserting that any of the things  
below actually happen in your environment. I am just listing some things  
to check for.

	1) You are not measuring sustained peak performance. Squid can go faster,  
and at higher loads you will see the dependency on the disk space  
utilization.

	2) You are overloading Squid and the bottleneck is elsewhere at these  
loads. Disk utilization matters, but not under these overstressed  
condition.

	3) You are not using enough of the disk space. There is always enough  
free disk space for filesystem performance to remain unchanged even at  
100% cache utilization.

> In addition to this I set low recurrence in a robot
> section. It seems to me logical that lower hit ratio then greater number
> of disk I/O operations. Perhaps hit ratio should be equal to zero.

In most cases, you want the workload to mimic your real environment. Use  
hit ratio close to what your environment has. Otherwise, you will be  
getting performance numbers that are of little use to you (whatever those  
numbers are). Also, by removing hits, you are reducing disk read  
operations in favor of disk writes. In some cases, it might make  
filesystem work easier, not more difficult.

HTH,

Alex.