Re: some linux tuning (fs benchmarks)

On Mon, 28 Jun 1999, Doug Renner wrote:

> Overall, it's only cache speed that we care about, right?

Iff your squid is only going to be operating from cache. Most people
running medium to large squid servers don't have enough RAM for that to be
the case.

> If I have maximum_object_size set to 4K, that's going to be the
> majority of my I/O. I will want to duplicate my environment as much as
> possible.

Do you really use maximum_object_size 4K? That's 3 orders of magnitude
smaller than the default (4MB). But lets assume you do have it set to 4K.
If you want an accurate test, you would create small files (many of them,
all together occupying several hundred MB to several GB) in the directory
structure squid uses, create a list of all these filenames in memory, and
then start randomly accessing files from the list (reading and
occasionally (with some set frequency) writing new data to them). You'd
want to measure how many accesses/second you can do. i.e. write a loop
that randomly selects files from the list, and every N reads, rewrite a
file with new data. Run through this loop X times (where X is probably
some number in the thousands) and time how long the sequence takes.

This would be far more meaningful a test than running bonnie a bunch of
times serially with smaller than physical memory file sizes.

> a better methodology. However, the real root of the thread was based on 1
> vendors card being faster or slower than the "md" driver in Linux. From
> what I understand, the bake-offs won't have those bounds. That's what I'm
> really looking for: justification that I'm making the right platform

If you really just want to compare cards, I think you can get at least
somewhat meaningful results with bonnie and a large file size. This
should give you an idea of the max sequential troughput of the
controller/disks, as well as an idea of how many seeks per second the
combination can handle. The previously mentioned test suite could be used
however to compare totally different hardware, filesystems, operating
systems, etc.

----don't waste your cpu, crack rc5...www.distributed.net team enzo---
 Jon Lewis *jlewis@lewis.org*| Spammers will be winnuked or
 System Administrator | nestea'd...whatever it takes
 Atlantic Net | to get the job done.
_________http://www.lewis.org/~jlewis/pgp for PGP public key__________
Received on Mon Jun 28 1999 - 22:03:09 MDT