Simon Fischer writes:
> I ran some tests with the 'dbench' benchmark from the samba project which
> simulates the disk i/o of the 'netbench' benchmark.
>
> I use 2.4.16 with the acl-patches (ext2/3) on a 128MB RAM system under zVM
> and initially tested dbench simulating 10 clients with ext2, ext3 and
> reiserfs on a minidisk. We have a G5(single proc) server connected to our ESS
> (shark) with 4 ESCON Channels (...and I think our machine is only lightly
> loaded).
>
> With ext2 I got 20-23 MB/s in several runs and with ext3/reiserfs 10-13 MB/s.
> Now I wonder what the bottlenecks are supposed to be and what to do to tune
> this before I'll make some further testings?! RAM & CPU seem to be OK for
> this workload and I thought 4 ESCONs should be much faster.

Others have already mentioned some of the issues. I'll just add a
couple of my own:
(1) dbench is an exceptionally poor benchmark in many ways--it not
only doesn't measure what it appears to be measuring but working out
exactly what it does stress/measure is very difficult. It also is
extremely sensitive to random environmental conditions: for example,
if the scheduler happens to run the multiple client processes in a
batch style then the overall throughput increases hugely (but
completely unfairly).
(2) One (of the many) potential constraint which might be affecting
your case is per-subchannel constraints. Instead of only using one
minidisk, try many (16, say, mounted on /vol/0, /vol/1, ..., /vol/f
with symlinks from the 128 CLIENTn directories used by dbench to
subdirectories of /vol/m allocated round robin). That way, Linux
will be able to queue up plenty of requests on each subchannel
(though that itself pre-supposes that dbench's behaviour isn't
throttling elsewhere depending on things like number of request queue
slots and whether I/O throttling in your particular kernel version is
done Linus-style on request slots or Alan-style elsewhere :-). It also
could easily just push the constraint further up the chain with no
gain (but I've definitely got higher figures than your running dbench
on a vaguely comparable S/390).

--Malcolm

--
Malcolm Beattie <[EMAIL PROTECTED]>
Linux Technical Consultant
IBM EMEA Enterprise Server Group...
...from home, speaking only for myself

Reply via email to