SMP - Quad vs Dual core

[Agent]

Anyone know how much of a difference there is?

Cheers

[directhex]

the "smp" f@h client is hard-coded to 4-way

if you run it on a dual-core box, then you have 2 processes per cpu, which means the cpu has to spend time swapping contexts in and out

[Agent]

I'm running dual core here, but there is definitely 4 processes running?

4* FahCore_a1.exe

[directhex]

jms      27018 18188  0 06:47 pts/0    00:00:00 ./mpiexec -np 4 -host 127.0.0.1
jms      27019 27018 62 06:47 pts/0    08:57:59 ./FahCore_a1.exe -dir work/ -suf
jms      27020 27018 56 06:47 pts/0    08:06:52 ./FahCore_a1.exe -dir work/ -suf
jms      27021 27018 39 06:47 pts/0    05:35:10 ./FahCore_a1.exe -dir work/ -suf
jms      27022 27018 25 06:47 pts/0    03:40:37 ./FahCore_a1.exe -dir work/ -suf

[Agent]

ack, I didn't read what you actually wrote!

I understand it'll be swapping things about, but by how much, and whats that translate into performance hit wise?

[directhex]

ack, I didn't read what you actually wrote!

I understand it'll be swapping things about, but by how much, and whats that translate into performance hit wise?

depends on your OS kernel and CPU design

first, there's a question of timer resolution - i.e. how long a particular process has to operate before the OS can consider doing other stuff. on linux, this is something determined at compile-time in the kernel. on this system, it's 250Hz (i.e. things can run for 250 CPU cycles before the OS might swap something else in).

next, there's the question of how the OS deals with the competing processes - how fairly it allocates time between the f@h processes (and other processes requiring cpu attention like the mouse driver). this is all a bit up in the air, but certainly the 2.6 linux kernel was a phenomenal improvement over 2.4 in this regard

finally there's the question of CPU design. swapping out one process for another takes a bigger performance hit on some CPUs than others - for example, remember Prescott P4? the big problem there was a massive pipeline - it took over 30 cycles for the cpu to actually start outputting work, after every context switch. multiple processes, more switching, more impact sue to the long pipeline

[Agent]

I'm aware of the majority of that, although Kernel level stuff is over my head, but I was looking for actual figures / benchmarks in the field

I'm just trying to justify buying a quad

[kidzer]

You will get more F@H points, Quad Core Justified

[Fraz]

I'm just trying to justify buying a quad

It's twice as good.

Will that do?!

[TheAnimus]

Kernels have really come a long way since pre-emptive multithreading, which iirc was in windows in NT4 but took just linux a little bit longer.

There are so many issues that will bottle neck the domestic multi core pc. Resources like an IDE CD Drive can 'lock' all the cores. This is a very common performance killer.

Also sorry hex, the EE inside me can't let you use 250hz like that... you explain 250 cycles, which of course, might not occur 250 cycles a second.

[directhex]

Also sorry hex, the EE inside me can't let you use 250hz like that... you explain 250 cycles, which of course, might not occur 250 cycles a second.

you're right, my mistake. i think it's indeed 250Hz, regardless of clock speed

[TheAnimus]

I must admit as of 2.6 i thought it was ment to be adaptive? (kinda like NT has been well, since MS bought the code)

[directhex]

2.6 made it configurable (at compile time), previously it was hard-coded to 1KHz

a very recent patch brought in a "tickless" kernel which removes interrupt polling entirely, though i forget which version that is (and it's not the default in most distros) (edit: 2.6.21)

[aidanjt]

I believe the tickless patch was only included in vanilla 2.6.23.

[joshwa]

My quad core (2.4 @ 3.0ghz) takes just over 15 hours to complete a work unit (500,000 steps)... my dual core (2ghz) takes roughly 41 hours* to complete the same amount of work...

* based on 1 pc taking 25 minutes.

[Agent]

I'm on a core 2 duo @ about 3.0ghz, takes about 30 hours (ish), going on the last 20% taking 6 hours.
So that is around double then Josh - Interesting, seems to be highly optimised for quads then

Think I will do a bit of saving, get the new water cooling kit first, so I can clock it like mad