Squid Cache Partition Performance Comparison of Reiserfs and Ext2

Squid Cache Partition Performance Comparison of Reiserfs and Ext2
Workload Simulated by Polygraph 2.2.7 polymix-1.pg

Squid Proxy Machine
Pentium III 450mhz, Abit BE6 Motherboard, IBM 20.5GB 7200rpm UltraATA-66, 128MB RAM
Redhat Linux 6.1, 2.2.13 Linux Kernel w/ Hedrick's IDE patch (HPT366) and linux-2.2.13-reiserfs-3.5.13-BETA-online-resizer-support-for-old-formats-plus-bugfix.gz
SQUID-2.2.STABLE5

Polygraph Client
Pentium 200mhz, 64MB RAM, Redhat 6.1

Polygraph Server
Pentium 200mhz, 64MB RAM, Redhat 6.1

 -- ReiserFS Results --
Full Logs - http://www.togami.com/laven/archive/1999/reiser-ext2/reiser.log.gz
Full lx Output - http://www.togami.com/laven/archive/1999/reiser-ext2/reiser.lx

req_rate: 47.83
rep_rate: 47.80
conn_ttl_mean: 18812.52
conn_use_mean: 1.00
req_dhr: 55.07
req_bhr: 66.77
rep_dhr: 53.29
rep_bhr: 64.57
rep_dcr: 80.31
rep_bcr: 97.40
rptm_mean: 18812.52
rptm_mean:hit: 18219.75
rptm_mean:miss: 19488.74
rep_sz_mean: 13693
fill_sz: 745104768.00
fill_cnt: 71739
err_cnt: 8272
err_pcnt: 5.11
duration: 3240.00

 -- ext2fs Results --
Full Logs - http://www.togami.com/laven/archive/1999/reiser-ext2/ext2.log.gz
Full lx Output - http://www.togami.com/laven/archive/1999/reiser-ext2/ext2.lx

req_rate: 45.51
rep_rate: 45.44
conn_ttl_mean: 15294.33
conn_use_mean: 1.00
req_dhr: 55.09
req_bhr: 66.84
rep_dhr: 52.40
rep_bhr: 63.31
rep_dcr: 80.28
rep_bcr: 97.40
rptm_mean: 15294.33
rptm_mean:hit: 13994.46
rptm_mean:miss: 16725.03
rep_sz_mean: 13727
fill_sz: 735871540.00
fill_cnt: 69560
err_cnt: 15627
err_pcnt: 9.66
duration: 3240.01

My Novice Interpretation
----------------------
The err_pcnt of the ext2fs test is nearly double that of the reiserfs test, but almost all of these errors are the result of "Too many open files" errors on the client machine. I'm guessing that this overflow of file descriptors on the Polygraph Client machine is indirectly related to the filesystem on the squid machine, as these errors are the result of piled up requests which are unfulfilled due to the constant overload on the squid box. The reiserfs tests achived a higher average fulfilled requests per second than ext2.

Conclusion
----------
Reiserfs appears to be more efficient (faster) for squid cache partitions than ext2.

Flaws in Experiment, Possible Areas of Improvement
-----------------------------------------------------
This is the first time I am using squid and polygraph, so any help in data interpretation would be appreciated. I do admit that my testing is flawed in many respects, but these tests were rushed and I don't have much hardware available. I'm in a situation where funds are extremely limited and high performance squid proxy caching is needed, so I'm not using SCSI disks and I'm trying to eke out any extra performance with reiserfs instead of the standard ext2. Does anyone out there have any experience with either reiserfs or the HPT366 controller, as they are both very non-standard and their stability is questionable. If anyone knows of any issues with either system element, especially in squid utilization feedback will be greatly appreciated.

Perhaps if this test, results and analysis were cleaned up, does anyone know if this type of thing could be included in the www.squid-cache.org benchmarking section?

Warren Togami
warren@togami.com
http://www.togami.com
Received on Mon Dec 27 1999 - 02:52:23 MST