This is great to see. I've put a Linux-port section on the main FileBench
wiki page.
We should try and get this integrated backinto a new public source repo.
Current one is on sourceforge, but is very out of date.
Do you know what Sun version you forked from? Does it have all the latest
fixes th
Try changing the run 10 to sleep 10...
Richard.
On 12/2/08 8:34 AM, "Demetri S. Mouratis" <[EMAIL PROTECTED]> wrote:
Gang,
I'm having trouble getting the 'foreach' syntax to loop through different I/O
sizes. Here is a sample of code I'm trying to get working:
#!/opt/filebench/bin/go_filebe
Drew: I had sent Erik a wad of changes that added a "raw:" syntax to enable raw
devices in the path string. Do you know if that was integrated?
Thanks,
Richard.
> -Original Message-
> From: [EMAIL PROTECTED] [mailto:perf-discuss-
> [EMAIL PROTECTED] On Behalf Of Andrew Wilson
> Sent: M
I'm almost certain that nobody is compiling recent versions on Solaris 8. The
makefiles did work on s8 at some stage, but the more recent changes have been
targetted at s10/11...
Richard.
On 7/8/08 1:28 AM, "Demetri S. Mouratis" <[EMAIL PROTECTED]> wrote:
Has anyone successfully compiled fil
Which OS is this happening on Doug?
On 6/4/08 12:05 PM, "Doug" <[EMAIL PROTECTED]> wrote:
> I am having a strange issue with filebench (which I think is caused by the
> profile).
> When I try to load a profile to run some checks in automated fashion, nothing
> seems to happen
> The output I ge
>>
>> OLTP
>>
>
> Isn't this the case of an application (e.g. vdbench, filebench) having
> slightly differing behavior and the underlying storage being a better
> match to one or the other? I am not surprised by this. If you were
> to say the workload was something like sequential read, then
FYI - I've made a patch for this as part of my current set of raw-disk
changes...
On Apr 28, 2008, at 0:17, [EMAIL PROTECTED] wrote:
> Hi,
>
> while running several storage performance tests using filebench
> (OLTP workload), it occurs to me that it would be good to have
> ability to put "red
I had to add quite a bit of code to make the database (oltp) workload run
properly on Linux, around the way that I/O is submitted (using Oracle's
io_submit instead of aio_write), and the way shared memory is allocated and
used. I also added a feature to enable raw: syntax in the filenames, so that
We have a great chapter in solaris performance and tools about
programming kstat...
Richard
On Apr 19, 2008, at 18:42, "Rayson Ho" <[EMAIL PROTECTED]> wrote:
> Have you read:
>
> http://developers.sun.com/solaris/articles/kstatc.html
> http://developers.sun.com/solaris/articles/kstat_part2.h
Grab the 1.0-linux tarball -- I don't think all the Linux changes have
been cross-ported into 1.1 yet...
Richard
On Mon, Mar 10, 2008 at 06:43:13AM -0700, Christian Fritze wrote:
> Hello everybody!
>
> I would like to try filebench on SLES 10 SP1 (2.6.16.46-0.12-smp SMP x86_64)
> and
> Debian
Indeed, the chimney seems to be bountiful -- We're hoping to use the
Linux builds over the break, your timing is most excellent!
Thanks for the update
Richard
On Wed, Dec 19, 2007 at 09:19:12AM -0800, eric kustarz wrote:
> I'm happy to report that we have updated the building process on the
>
This could be a bug: the actual file sizes in a set are a gamma
distribution, with the median at $filesize. Using a file should
yield something different. Perhaps some of the logic that sets file
sizes in light of a singular set needs looking at?
On Mon, Oct 29, 2007 at 01:14:28PM -0700, John Vi
rd.
>
> Regards,
> John G
>
> - Original Message
> From: Richard McDougall <[EMAIL PROTECTED]>
> To: John Vijoe George <[EMAIL PROTECTED]>
> Cc: perf-discuss@opensolaris.org
> Sent: Friday, October 19, 2007 11:23:16 AM
> Subject: Re: [pe
Great!
Yes, there was a known issue with some workload - doing repeat runs
may require exiting and calling into the interpreter again.
I believe this is fixed in the new source release, which is being tidied
up for Linux...
Richard.
On Fri, Oct 19, 2007 at 07:06:03AM -0700, John Vijoe George w
You might need to install filebench explicitly in /usr/local/filebench
on Linux to avoid this error...
Richard.
On Thu, Oct 18, 2007 at 03:13:41PM -0700, John Vijoe George wrote:
> Hi,
> I am trying to get filebench to run on my Linux ( RHEL 5 ) machine -
> 2.6.18-8.el5. I managed to install f
Check out mdb's ::memstat command
Fire up mdb -k, then type in ::memstat.
The pagecache and cachelist are the two components of the file cache.
You may find some of this documented in the book and on the wiki
(solarisinternals.com).
Richard.
On Sep 21, 2007, at 10:40 PM, Alexandr Semenov wr
Hi Bob,
> I've often wanted to see prstat changed to consistently show 'percent of
> a CPU core',
> so that values might exceed 100 for multi-threaded apps, and so
> per-process values
> from the same task running on different configurations would be more
> comparable.
Can you expand on the sc
Hi Cherian,
The -m shows microstate percentages for each process; i.e. the percentage of
time this thread spent on cpu as a percentage of the elapsed time of the
sample. It doesn't correspond to the regular prstat percentage of the whole
machine value - so it's actually a different metric.
W
Setting larger maxphys has in my experience made quite a difference,
depending on the target. (actually on x64 it makes it
slower). Generally, performance increases as the transfer size
increases has been my observation, for some classes of targets. This
is likely due to the target's ability to se
Hi Roman,
Could some of this be file system related? Your observation about make running
slow hints
to many areas, some of which could be the default file system config.
It would be interesting to run your test(s) in /tmp as part of a
rudimentary investigation into the problem area...
Creating
Hi Leon,
You should be able to get 4x that level of throughput on 2 opteron
processors. Some other tuning that might be useful:
You might increase the number of servers higher (there isn't any
significant downside). We use 1024 as a default on most systems.
In /etc/system:
set ip:ipcl_con
Yes; a simple change the the workload definition will create threads
vs processes.
The clause:
define process, instances=n
{
thread , instances=n
{
...
}
}
You can arrange the process/thread relationship anyway you like. In
the simplest case to make some
Also, here are similar scripts for the cpu memory hierarchy.
Richard.
On Thu, Nov 03, 2005 at 11:16:25AM -0500, Matty wrote:
> On Wed, 2 Nov 2005, Richard McDougall wrote:
>
> >
> >Hi Matty,
> >
> >We use some scripts to post-process busstat into what you are look
IN a nutshell yes, but...
THe default for filebench today doesn't produce a throughput summary
at regular intervals during the run. It would be quite easy to add
this feature, which is something Roch has asked for.
It's possible to do this by extending the commands sent into the
interpreter. Som
AbUtil = Address bus util
DBUtil = Data bus util
RTS->WS are MOESI bus transactions (request to send, etc...) from the
PRM.
Thanks,
Richard.
On Thu, Nov 03, 2005 at 11:16:25AM -0500, Matty wrote:
> On Wed, 2 Nov 2005, Richard McDougall wrote:
>
> >
> >Hi Matty,
> >
Hi Matty,
We use some scripts to post-process busstat into what you are looking
for: bandwidth per sbus and device. It's a two phase operation now,
collect and report. See below for the report and collect scripts...
Regards,
Richard.
On Wed, Nov 02, 2005 at 01:48:49PM -0500, Matty wrote:
Hi Tim,
At this time, the NFS plug in is on the to-do list. We are looking at
taking the NFS4shell (on sourceforge) and using it to generate NFS4
protocol requests from userland.
Today, there are a few minor changes required to make the plugin; 99%
of all I/O operations are in the flowop_library
Not at the moment; Filebench uses /dev/random, rather than a
repeatable random sequence...
Thanks,
Richard.
On Thu, Oct 13, 2005 at 04:47:50PM -0700, eric kustarz wrote:
> Is there a way to set the seed value for random IO workloads in filebench (so
> that i can fairly compare random IO runs)
> So if i understand this correctly, it will create $nthreads, and each of
> those threads will create two threads: one to do "rand-read" and one to
> do "rand-write". And there's no serialization between the "rand-read"
> and "rand-write" threads right (that is, the rand-write doesn't have t
Hi Matty,
Kernel async I/O just works on RAW devices at the moment.
Kernel async I/O give a small incremental improvement, but only in the
order of a few percent in optimal cases. The user-land thread
emulation works quite well when direct I/O is enabled (without it, the
single writer lock will
30 matches
Mail list logo
