File:Gmirror-performance.png

Revision as of 00:09, 27 December 2007

Chart showing performance of gmirror (and some proprietary RAID1 hardware implementations).

System: 
   amd64 6.2-RELEASE, Athlon X2 5000+, 2GB RAM
   3x Western Digital 250GB drives (WDC WD2500JS-22NCB1 10.02E02 SATA-300)
   2x Western Digital 500GB drives (WDC WD5000AAKS-00YGA0 12.01C02)
   Nvidia nForce onboard RAID controller, Promise TX2300 RAID controller

Each process shown in the graph is a cp to /dev/null of an individually generated 3200MB random binary file (using dd if=/dev/random bs=16m count=200).

The "split" balance algorithm was also tested but resolved in universally horrible performance, so was not included on the graph.

For these tests, read-ahead cache value was GREATLY increased from 6.2-RELEASE's default of 8: sysctl -w vfs.read_max=128.

Raw data follows:

Nvidia onboard RAID1: 2x250		60	108	90	96	100	* 15 secs delta on 2-process copy / 38 seconds on 5-process copy!
Gmirror: 3x250/2x500 round-robin	63	73	69	71	72		
Baseline: 1x500			76	70	72	72	72		
Baseline: 1x250			56	56	55	55	56		
Gmirror: 3x250 round-robin		59	53	53	56	55		
Promise TX-2300 RAID1: 2x250		61	52	54	55	57	* 15 sec delta on 5-process copy
Gmirror: 3x250/2x500 split 128k	54	50	47	47	47		
Gmirror: 2x250 round-robin		58	49	51	51	50		
Gmirror: 2x250 split 128k		55	47	47	47	48		
Gmirror: 3x250 split 128k		55	46	49	55	54		

"Delta" in raw numbers refers to individual processes finishing significantly faster than others, though all were started simultaneously.