Dear All,10 days ago I have reported a problem with OpenAFS on Linux to openafs-bugs (#133467). Now I have found more results and a solution to circumvent this problem.
I have a simple program executing the following calls again and again: fopen afs-file mode append fwrite 64MiB fclose The machine has a disk cache with 16GiB and afsd is started with -afsdb -blocks 14862564 -stat 65536 -volumes 128 -chunksize 18 The first 40 64MiB writes took about 0.7s each, then the time toappend 64MiB is 1.2s for block #41 and is growing to 11s for write #154, then I run a "fs flushall" and the write time drops to 0.8s, but is growing again. When I kill the process, I get a kernel crash most of the time.
When I use afsd -dcache 32768, I see the following: for the first 128 writes (=128*64MiB=32768*chunksize) the time for each write is growing from 0.7s to 1.5s, then it jumps to 6.5s and is growing to ~10s for write #154 and ~19s for write #220
For our production machines I have set the chunksize to 2^21 and use the default dcache size (10000?). Now we do not see any problems with our current jobs.
It seems that the algorithm to look for free dcache entries is very slow. Any hints?
Best regards, Hans-Werner
smime.p7s
Description: S/MIME Cryptographic Signature
