Path: utzoo!utgpu!jarvis.csri.toronto.edu!rutgers!usc!cs.utexas.edu!uunet!zephyr.ens.tek.com!tektronix!sequent!venkat
From: venkat@sequent.UUCP (Venkat Malla)
Newsgroups: comp.protocols.nfs
Subject: Re: Nhfsstones
Keywords: nhfsstones, filesystems
Message-ID: <27392@sequent.UUCP>
Date: 9 Jan 90 03:52:20 GMT
References: <3392@rti.UUCP>
Reply-To: venkat@sequent.UUCP (Venkat Malla)
Distribution: comp
Organization: Sequent Computer Systems, Inc
Lines: 35

In article <3392@rti.UUCP> mcm@rti.UUCP (Mike Mitchell) writes:

>I've been playing with the 'nhfsstone' program from Legato Systems, Inc.,
>and I've been wondering a few things.  As it comes from Legato, 'nhfsstone'
>can only check out NFS file systems.  I looked at the code, and it reads a
>kernel data structure to get statistics.  The statistics that is uses are
>operation counts.  Each sub-process knows how many operations that it has
>performed, but the parent process does not know how many operations each
>child has performed.  Why can't the children communicate this information
>to the parent using a temporary file or pipes?  Why is reading the kernel
>data structure necessary?  (Because it is there?)

What the sub-processes do is generate pseudo NFS ops to simulate an
NFS environment on the client side.  Since these dont necessarily 
reflect in the actual NFS ops performed, the kernel NFS stats are read
periodically and the operation-mix is monitored (couldn't figure out the
formulas used here, anybody did?) and the NFS load adjusted.  

>
>I modified the program to use shared memory to pass the data around, and
>now I can get statistics for local file systems.  Does anyone see any problem
>with this approach?  

The parent merely sets up the client processes and finally gathers
the stats they record.  You dont need to pass the data around.  
Each sub-process is independent and acts like a separate client (to   
simulate several processes grabbing the few file-handles available).

>
>I would like a way to compare other filesystems' performance to NFS performance
>in a semi-standard way.  Does anyone have a better approach?
>
Nhfsstone is tailored to measure NFS performance (diff. implementations).
Other file systems may use a different protocol.  You could perhaps use
the connectathon general tests for such comparison.