These are the afs over atm benchmarks we promised. I ran these tests
about a month ago using a very early release of afs 3.4 on a few powerful
aix systems. The program I used to benchmark afs over atm was a modified
version of bonnie. I was only able to use two machines for the benchmark.
We hope to try n clients to 1 server tests, where n is greater than 1, in
the very near future.
Below is a profile of the client and server machines as well as the
benchmark software.
client:
rs6k 590 512MB of RAM (sp2 wide node)
one 2 gig scsi II disk (standard IBM)
aix 3.2.5
very early release of afs 3.4
server:
rs6k 550 128MB of RAM
one 2 gig scsi disk (unknown origin)
aix 3.2.5
very early release of afs 3.4
benchmark software:
bonnie (Author Tim Bray)
source can be obtained via archie
Sequential Output
Per-Character: using putc(1) to write file
Block: using write(2) to write file
Rewrite: using read(2), write(2) and lseek(2)
Sequential Input
Per-Character: using getc(1) to read file
Block: using read(2) to read file
Random Seeks
4000 lseek()'s in file determined by random()
For the cold cache definitions below I modified bonnie to do an "fs
flushvolume" to force the volume out of disk cache and clear out the
rs6k's memory buffer.
The following results are based on a cachesize of 10 megs and a chunksize
of 1 meg using a very early release of afs 3.4. The MB column represents
the size of the file we benchmarked in megabytes. If the %CPU is above 30%
I think we can safely assume that afs is taking advantage of the rs6k's
memory to disk caching subsystem.
-------Sequential Output-------- ---Sequential Input-- --Random--
-Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks-
Cache MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU
cold 20 901 59.5 2195 8.9 872 4.2 944 62.1 2147 3.9 56.2 1.2
warm 20 1307 41.0 2180 9.5 1821 10.2 3281 99.2 119510 99.2 118.7 1.5
cold 10 1002 68.8 3039 12.5 811 4.6 947 62.1 2266 4.2 99.4 1.9
warm 10 1622 49.8 3091 10.9 1856 8.3 3277 100.2 119355 93.2 164.7 1.9
cold 5 1376 90.8 12306 55.3 851 4.2 946 62.0 2249 4.0 162.7 2.9
warm 5 2606 86.5 13107 48.6 1818 9.6 3280 100.0 118274 92.4 1257.2 14.8
The following results are based on a disk cachesize of 10 megs and a
chunksize of 64k again using a very early release of afs 3.4
-------Sequential Output-------- ---Sequential Input-- --Random--
-Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
Cache MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU
warm 20 894 27.6 957 4.4 499 2.7 3272 99.9 115565 95.9 120.0 1.9
warm 10 1113 33.4 1343 5.6 516 2.1 3274 99.8 114583 89.5 206.1 3.3
warm 5 1317 39.4 1892 7.8 1689 6.3 3269 99.0 114089 111.4 1239.2 14.6
I am sorry I do not have any afs 3.3a results. We found out about the MTU
problem on multi homed hosts a while ago and decided to go with an very early
release of 3.4 which did select the correct MTU for atm.
I think it is safe to say that "very early" means pre beta.
Thanks,
Lou
Louis R. Constable
Systems Programmer
Cornell Theory Center
Ithaca, NY 14850
