For these multi-core processors, I'd be interested in benchmarks that exercise the system while doing more than one thing at a time.
Synthetic benchmarks are all well and good, but how does the system do when trying to *really* multitask? Consider a server (which is where the Xeon is really targeted IMHO), how does it look when getting concurrent loads (from a network, not the local machine) on Apache, MySQL, and PHP?
I know there are tools for artificially loading servers ('siege' is an http tool that comes to mind), but I do not know if there is a benchmark "suite" that makes the process of setting up and running the tests manageable.
Also of note is the fact that the 2.6 kernel has tuning parameters for a lot things that can really affect overall system performance. Perhaps running the tests with kernels configured for "desktop" and "server" tuning sets would be interesting. (Please note, I am not a kernel hacker.)
Kernel parameters that I think could be considered include:
- Processor Family (e.g. - *not* using a generic selection)
- Max number of CPU's
- SMT and Multi-core (CPU) schedulers
- I/O schedulers (Anticipatory, Deadline, CFQ)
- Preemption Model (No Forced (Server), Voluntary (Desktop), Preemptible (Low-Latency Desktop))
- Preempt the Big Kernel Lock (Desktop)
- High Memory Support (select "4GB" for between 1 & 4GB RAM!)
- Allocate 3rd-level pagetables from highmem
- MTRR support (not selecting this will cause graphics support to *really* suck.)
- Enable kernel IRQ balancing
- Use register arguments
- Timer frequency (lower for servers, higher for desktops)
- Choice of on-disk filesystems (ext2, ext3, reiserfs, xfs, etc.)
Oh, and don't forget all of the kernel "patchsets" out there that further tweak the kernel.
Thanks for listening.



Reply With Quote
(joke, but close). 
