On Wed, 18 Feb 1998, Lyle Seaman wrote:
> At 11:59 AM 2/18/98 +0100, Hartmut Reuter wrote:
>
> >I cannot really see what advantage the load balancing across spindles and
> >SCSI buses should have since all I/O is synchronous. There will never be
> >more than one I/O request active at a time. The only improvement you can
> >get is that from disk striping which shortens the time needed for a
> >single I/O.
>
> It's not _completely_ synchronous. Writes are almost entirely asynchronous,
> as are prefetches. The file server reads data in 8 KB buffers, so if you're
> using the default 64 KB chunk size, the first read is synchronous, and as
> many as seven subsequent reads may be asynchronous. Then, there's a final
> prefetched page that sits in the buffer cache, unused, though a subsequent
> FetchData RPC may pick it up before it is flushed.
What you say about the write() calls is certainly true. For the read()
case things are different in MR-AFS: here the buffer size is eather
64 KB (for small files) or 1 MB (for big files). Therefore you can gain
much performance by striping the partitions.
I suppose the large buffer sizes have been chosen to minimize
protocoll overhead:
In MR-AFS all I/O goes through a protocoll layer which decides whether
the data can be read or written locally or have to be transferred over the
network from/to a remote residency (e.g. an archive server).
The data rates we see writing and reading files from a client are never
the less not bad
3 to 4 MB/s for write and
4.5 to 5 MB/s for read
with a file size of 10 MB,
client and server IBM rs 6000 on FDDI,
data stored on local SSA disks at the fileserver.
Would be interesting to get performance values for standard AFS in a
similar situation.
Another interesting fact is that during the last years the performance
has improved by at least 30 % probably due to the better rx-libraries
compared to AFS 3.3a.
-Hartmut
-----------------------------------------------------------------
Hartmut Reuter e-mail [EMAIL PROTECTED]
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)
-----------------------------------------------------------------
