zfs discuss - Jul 2008 - Measuring ZFS performance - IOPS and throughput

Can anybody tell me how to measure the raw performance of a new system
I''m putting together?  I''d like to know what it''s
capable of in terms of IOPS and raw throughput to the disks.

I''ve seen Richard''s raidoptimiser program, but I''ve
only seen results for random read iops performance, and I''m
particularly interested in write performance.  That''s because the live
server will be fitted with 512MB of nvram for the ZIL, and I''d like to
see what effect that actually has.

The disk system will be serving NFS to VMware to act as the datastore for a
number of virtual machines.  I plan to benchmark the individual machines to see
what kind of load they put on the server, but I need the raw figures from the
disk to get an idea of how many machines I can serve before I need to start
thinking bigger.

I''d also like to know if there''s any easy way to see the
current performance of the system once it''s in use?  I know VMware has
performance monitoring built into the console, but I''d prefer to take
figures directly off the storage server if possible.

thanks,

Ross
 
 
This message posted from opensolaris.org

Ross wrote:
> Can anybody tell me how to measure the raw performance of a new system
I''m putting together?  I''d like to know what it''s
capable of in terms of IOPS and raw throughput to the disks.
>
> I''ve seen Richard''s raidoptimiser program, but
I''ve only seen results for random read iops performance, and
I''m particularly interested in write performance.  That''s
because the live server will be fitted with 512MB of nvram for the ZIL, and
I''d like to see what effect that actually has.
>   
Cool.  Yes, RAIDoptimizer''s performance model is trivially simple
because
it uses disk datasheet specifications, not measured data.  There is a lot of
work being measured by filebench, which should be installed for you in
/usr/benchmarks
> The disk system will be serving NFS to VMware to act as the datastore for a
number of virtual machines.  I plan to benchmark the individual machines to see
what kind of load they put on the server, but I need the raw figures from the
disk to get an idea of how many machines I can serve before I need to start
thinking bigger.
>   
> I''d also like to know if there''s any easy way to see the
current performance of the system once it''s in use?  I know VMware has
performance monitoring built into the console, but I''d prefer to take
figures directly off the storage server if possible.
>   
Something like NFSstat is probably the best indicator from the client
perspective.
 -- richard

Ross Smith wrote:
> Thanks Richard, filebench sounds ideal for testing the abilities of 
> the server, far better than I expected to find actually.
>  
> NFSstat might be tricky however, since the clients are going to be 
> running XP :).  I''ve got a very basic free benchmark that
I''ll use to
> check that virtual disk performance over NFS is acceptible on the 
> client, and then I''ll use the performance figures on VMware and
the
> fileserver to see how the clients are doing once I''ve a few
running.
>  
> On the ZFS fileserver, is iostat all I need to get a quick snapshot of 
> the load on the system?  Is there anything on Solaris like
Microsoft''s
> performance monitor, where I can log figures over a period of time, or 
> bring up a chart of performance over time?  What I''d really like
to
> know is the average load in terms of iops and bandwidth, plus the peak 
> figures for each statistic too.
There is a performance monitor, actually man different ways to look at
performance.  In general, the best view is from the client.  The farther
you get from the client, the less you can see.  For example, using iostat
on the server is a common thing to do, but since the server caches
data from the disks and iostat only shows I/O  requests, it will be
difficult to correlate server disk I/O to client performance over time.
There is a wealth of information on Solaris performance analysis and
tuning knowledge available on the net.
 -- richard