Bruce Allen wrote:
spec.org has a benchmark that is nfs-specific but it does not give you
MB/s for a special access pattern.
Instead, it gives you operations per second as well as response time
at that load for a mix of different access patterns, filesizes between
1K and 1M.
http://www.spec.org/benchmarks.html#nfs
Can you tell me *roughly* how to translate this number into MB/sec? I
understand that it's a mixed load of requests with different sizes and
access patterns; still my question makes sense.
I can see that on an SGI system that *should* be limited by its network
connection (1 Gb/sec) the benchmark give around 13,000. So I would
guess that ~100 MB/sec corresponds to a SPECsfs97_R1.v3 (NFS version 3)
benchmark score of around 13,000.
Similarly, I see that on an IBM system with 12 gigabit ethernet ports,
the benchmark is around 118,000, which would correspond to around 900
MB/sec.
Michael, do these MB/s numbers look roughly right to you? In other
words that SPECsfs97_R1.v3 (NFS version 3) of 13,000 is around 100 MB/s?
Cheers,
Bruce
I don't think it is valid to try and convert the SPEC numbers to MB/s.
From the raw results, it appears that the SPEC numbers are only Ops/sec.
If each operation was to move a packet of data (1500 or 9000 bytes) you
could do the math. However, with file operations and everything else,
I don't think you can determine what you want from those numbers.
I think what this does help with is figuring out how the server can
handle load. The higher the ops/sec is, the more simultaneous clients
can be supported before performance starts to fall off. When I was
talking about my experiments with NFS server performance, I would run
lmdd from multiple clients. I was able to sustain the peak bandwidth
across 8 clients. As I added more clients, the total performance would
fall off. I did tweak NFS parameters (NFS threads, packet sizes, etc).
I just didn't have the horse power to support all of the connections.
If you really want to support alot of clients, then you need to worry
about your server configuration. It is nice that Hypertransport scales
so well. This is an application where dual core will be a big benefit.
Craig
_______________________________________________
Beowulf mailing list, [email protected]
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
