Subject: benchmark results + questions
From: Holger Kiehl (Holger.Kiehl@dwd.de)
Date: Fri Jan 14 2000 - 13:21:52 CET
Here are some results I got with bonnie:
-------Sequential Output-------- ---Sequential Input-- --Random--
-Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU
1 768 3328 46.2 3058 2.8 3601 7.9 7764 98.5 31810 21.1 284.4 3.6
2 768 6610 98.8 28498 38.2 12835 37.1 7323 93.7 28203 20.4 313.5 4.4
3 768 6525 96.9 40281 86.6 18871 57.6 7511 97.0 47529 37.2 344.8 5.3
4 768 6608 98.7 39775 63.1 17404 53.8 7463 96.0 41187 32.2 341.5 4.9
5 768 6582 97.6 38862 61.6 17510 51.0 7497 96.3 41476 30.5 346.0 5.5
6 768 6575 97.4 39471 60.1 15493 44.2 7577 97.3 40808 33.3 371.9 6.3
7 768 6625 97.6 40965 59.9 14219 40.3 7592 97.7 42571 35.9 271.1 4.9
8 768 7121 99.8 72177 73.5 21740 49.5 7711 97.9 64926 45.6 265.8 3.8
1) GDT 6518 RD (32MB EDO-RAM) Raid 5 (chunksize=? stride=8) 2.2.12
2) Software Raid 5 on one Adaptec U2W Controller (chunksize=64 stride=8)
One disk set to 10MB/s 2.2.12
3) Software Raid 5 on on Adaptec U2W Controller (4 disks) and Symbios 875
(2 disks) (chunksize=64 stride=16) 2.2.12
4) Software Raid 5 on one Adaptec U2W (chunksize=64 stride=16) 2.2.12
5) Software Raid 5 on one Adaptec U2W (chunksize=64 stride=16) 2.2.14
6) Software Raid 5 on one Adaptec U2W (chunksize=32 stride=8) 2.2.14
7) Software Raid 5 on one Adaptec U2W (chunksize=8 stride=2) 2.2.14
8) Software Raid 0 on one Adaptec U2W (chunksize=4) 2.2.14
2 PIII 450MHz on an Asus P2B-DS with 256 MB ECC RAM and 6 IBM DNES 9GB LVD
drives. Distribution is RH 6.1 and the block size of the file system is
always 4K.
Although the Symbios 875 is only an UW, it does not make much difference. In
fact I did some other benchmarks where lots of small files are being copied
and ftp'ed (locally) around (for 7 hours!), which showed the single controller
with better results.
What puzzles me is the bad results on the hardware raid controller when
using Bonnie. Doing the 7 hour benchmark with lots of small files I come
close to the results of a software raid system.
However, just writing a 30MB file to disk would produce the following
results on a hardware raid 5:
Write file ( 512Byte) : 0.365 s 84335 KB/s
With sync : 6.252 s 4913 KB/s
Write file (1024Byte) : 0.260 s 118260 KB/s
With sync : 5.935 s 5176 KB/s
Write file (4096Byte) : 0.182 s 169125 KB/s
With sync : 5.867 s 5236 KB/s
Write file (8192Byte) : 0.192 s 160496 KB/s
With sync : 5.872 s 5232 KB/s
and the same thing done on a software raid 5:
Write file ( 512Byte) : 0.465 s 66086 KB/s
With sync : 1.117 s 27521 KB/s
Write file (1024Byte) : 0.327 s 94183 KB/s
With sync : 1.20 s 30131 KB/s
Write file (4096Byte) : 0.322 s 95614 KB/s
With sync : 0.940 s 32699 KB/s
Write file (8192Byte) : 0.262 s 117377 KB/s
With sync : 0.947 s 32463 KB/s
[The "With sync" part is when putting an fsync() after the 30MB are written]
Why is the hardware raid so much slower when syncing to disk?
Holger
This archive was generated by hypermail 2b25 : Sat Jan 15 2000 - 17:48:02 CET