Samofatov, Nickolay wrote:

>Do you use NUMA-enabled 2.6 kernel and does your hardware/BIOS report
>node distances correctly?
>

I don't know, but I assume I can rebuild the kernal if not. Anything in
the interest of science.

>BIOS may not report NUMA support and interleave pages on different
>controllers to level access timings. Send your dmesg to me please.
>

I'm using the MSI K8D motherboard, which is pretty much the standard
server dual processor board, available in both ATX and rack mount
configurations. If you unable to show a performance benefit for what is
absolute high end of commodity boxes it rather calls into question the
judgement that generally accepted programming standards should be thrown
out the window by the Firebird project.

>Also NUMA affinity should not show big benefits with 2-way machine
>because single 2-channel memory controller (and hyperthreading to access
>it) is usually fast enough for 2 CPUs. I may try to run some tests on
>more representative NUMA, such as SGI Altix machine or some Sun box, but
>it will take some time since I left interesting boxes in russian
>regions. :)
>
>Look at this article:
>http://home.arcor.de/efocht/sched/
>
>Machines mentioned in this article have pretty cool memory architecture
>this is why real performance difference for numa-aware and non
>numa-aware allocations is so small (25%).
>
>
>

We can get way more than 25% performance improvement by better
programming practices. Eliminating the frequent use of thread data, for
example, is probably three times this. Eliminating redundant request
compilations, reducing active memory by using shared requests and
eliminating redundant fetches and copies of metadata will give
efficiencies far beyond 25%. But you can't use any of this because of
your rejection of the principles of object oriented design.

The only way you can possible justify your ideas is through a
demonstrated performance gain. I've got the appropriate hardward, let's
run the tests.

--

Jim Starkey
Netfrastructure, Inc.
978 526-1376