[PVFS-users] terrible performance ... help!

[Kaveh Moallemi]

Hi Troy,

Thanks for the suggestions.  The code snip that I provided is part of
a MPI program ... I decided to use native PVFS calls hoping to improve
performance (the program is IO intensive).

> PVFS with a single iod and a single client is not likely to perform
> well; the whole point of the file system's design is to to have multiple
> servers and multiple clients to increase the amount of aggregate
> bandwidth available.

Currently there are 1 mgr, 1 iod and 5 clients ... I'm going to add 4
iod's as soon as I get some disks.  To make use of all compute
processors, I'm running my code with 8 MPI processes.


> In addition, it's not clear from the code snippet
> you posted what the size of pvfs_buffer is, but if it's less than 64kB
> (the default stripe size for a file on PVFS), you're even less likely to
> see good performance.

Yes, I am infact writting less than 64KB at a time ... good point,
I'll modify my code.


> What are you trying to accomplish here?  PVFS is *not* a general purpose
> file system, so it does some things (eg. parallel I/Os to/from multiple
> hosts with very large block sizes) very well and others (eg. I/O to/from
> a single host with very small block sizes) rather badly.

My question is, with the current setup, why does the non-pvfs version
of my program performe so much better (fprintf() vs. pvfs_write())?
Note that both versions work out of the same pvfs mounted directory.


Thank you,


../kaveh