[OpenAFS] OpenAFS speed

Hartmut Reuter reuter@rzg.mpg.de
Wed, 25 Jun 2003 10:25:34 +0200


Stephen Joyce schrieb:
> On Tue, 24 Jun 2003, Hartmut Reuter wrote:
> 
> 
>>In our Linux-cluster with Gbit Ethernet
>>and ramdisk cache I see transfer rates of > 25 MB/s into AFS.
> 
> 
> I've never seen anyone claim transfer rates anywhere close to this with any
> brand of AFS (except for the testing that umich did ~10 years ago bypassing
> the cm and using dual paths).
> 
> Did you hack the client or server or do anything special to see this speed?

We are using a client version of OpenAFS based on the CVS-version from 
October last year. This because we need large file support and we also 
have AIX 5.1 machines. But I think there is nothing special with respect 
to perforance.

The fileservers are MR-AFS servers, of course namei, with the old lwp 
threading. In this special case we do not use shared residencies, but 
it's exactly the same as in OpenAFS, except the large files support.

The client machine is an IBM blade center PC, double processor, with 
Redhat 8.0. The server is a blade center head node (because of the 
FC-interface needed for the IDE-RAID5), also double processor, with 
Redhat 7.3. Writing directly into the vicepa-partition allows for a 
transfer rate of about 80 MB/s. All blade center machines have Gbit 
ethernet to a common switch.

I have a special test script to test the read and write performance. The 
script starts a program which gets two parameters: file size and number 
of streams. It forks itself number of streams times and each copy looks 
for a file "raidfile.stream-number". If the file exists it reads the 
file, otherwise it creates and writes the file.

The script does 7 tests with the following numbers of streams: 1, 1, 2, 
4, 4, 8, 8

That means: 1st it writes a big file and measures performance, then it 
reads it. Then it writes another big file and reads at the same time the 
first one. Then it writes two more big files and reads the existing two 
files in parallel. Then the same with 4 new files ending up with the 
parallel read of 8 files. Here we come to the limit of 4 calls per 
connection

Because of the size of the files (1GB or 8GB) and the use of only 16 MB 
ramdisk  as cache AFS client caching cannot have any effect on the read 
performance.

Output of the script for 8 GB files is the following:

==========================================================

running on machine Linux bc-c4.rzg.mpg.de 2.4.18-27-compute #3 SMP Thu 
Apr 3 18:16:46 CEST 2003 i686 unknown
at Tue Jun 24 17:15:11 MEST 2003
in directory /afs/ipp/home/h/hwr/test/bc-hn2
Using files with option -g and size 8
1st test: write raidfile.0

real    5m7.033s
user    0m1.640s
sys     2m53.870s

2nd test: read raidfile.0

real    4m38.770s
user    0m2.680s
sys     0m39.280s

3rd test: write raidfile.1 and read raidfile.0

real    8m21.970s
user    0m4.370s
sys     3m46.940s

4th test: write raidfile.2 and raidfile.3 and read the others

real    16m24.759s
user    0m9.550s
sys     9m8.730s

5th test read 4 files in parallel

real    15m30.249s
user    0m10.640s
sys     7m12.800s

6th test: write raidfile.4 to raidfile.7 and read the others

real    33m34.037s
user    0m18.340s
sys     19m58.660s

7th test read 8 files in parallel

real    29m6.934s
user    0m22.340s
sys     16m6.000s

Average values:

write 1 stream      27338 KB/s
read  1 stream      30100 KB/s
write 2 streams     21809 KB/s
read  2 streams     16715 KB/s
write 4 streams     10554 KB/s
read  4 streams      8883 KB/s
write 8 streams      5795 KB/s
read  8 streams      4612 KB/s

=====================================================================

The number of streams for the average values is the number of copies of 
the program running at the same time. So "write 8 streams" is actually 
the average of the 4 write streams in test 6 while 4 read streams were 
running at the same time.

The aggregate transfer rate comes close to 40 MB/s in the case of 4 and 
8 streams parallel.

But still less than 1 CPU is used in average at the client and also at 
the server side and we are far from the network bandwidth. So there is 
still some unnecessary waiting.

If someone is interested in the script and program he/she can find them in

/afs/ipp-garching.mpg.de/u/hwr/public/performance-tests

-Hartmut
> 
> Cheers,
> Stephen
> --
> Stephen Joyce
> Systems Administrator                                            P A N I C
> Physics & Astronomy Department                         Physics & Astronomy
> University of North Carolina at Chapel Hill         Network Infrastructure
> voice: (919) 962-7214                                        and Computing
> fax: (919) 962-0480                               http://www.panic.unc.edu
> 
> "UNIX *is* user friendly. It's just selective about who its friends are."
> 
> _______________________________________________
> OpenAFS-info mailing list
> OpenAFS-info@openafs.org
> https://lists.openafs.org/mailman/listinfo/openafs-info


-- 
-----------------------------------------------------------------
Hartmut Reuter                           e-mail reuter@rzg.mpg.de
					   phone +49-89-3299-1328
RZG (Rechenzentrum Garching)               fax   +49-89-3299-1301
Computing Center of the Max-Planck-Gesellschaft (MPG) and the
Institut fuer Plasmaphysik (IPP)
-----------------------------------------------------------------