A thought regarding max output per cpu, HT style

[aberant]

Its my understanding that the basis of Hyper Threading is that not all parts of the cpu are being utilised at any one time, i.e. the FPU might be doing something whilst the ALU just sits there. So, is it possible to identify projects that rely almost solely on the FPU and another that relies almost slowly on the ALU ? That way we could run both projects at once. Given, one might lose some output if there were one sole project running, but combined would it not give a greater output ? Obviously there's memory bandwidth to consider as well, and i'm no Computer Science boffin, so perhaps someone could enlighten me (BK would be good right about now ! )

 

[Assimilator1]

ttt

(no idea in other words,though I do know that members who run HT rigs with 2 instances of SETI get a significant output boost)
Btw you might want to include HT in your title to grab more attention

 

[Unforgiven]

i dont have any clue what hes talking about but i have 2 hyperthreading rigs (2.6c & 3.0c) that crank out 15-16 units a day respectively.

 

[ProviaFan]

In theory, that should work, but I'm no hardware design guru so I'm not the one to ask.

 

[GunDog]

Originally posted by: aberant
Its my understanding that the basis of Hyper Threading is that not all parts of the cpu are being utilised at any one time, i.e. the FPU might be doing something whilst the ALU just sits there. So, is it possible to identify projects that rely almost solely on the FPU and another that relies almost slowly on the ALU ? That way we could run both projects at once. Given, one might lose some output if there were one sole project running, but combined would it not give a greater output ? Obviously there's memory bandwidth to consider as well, and i'm no Computer Science boffin, so perhaps someone could enlighten me (BK would be good right about now ! )

The biggest problem is finding a DC Project that uses just FPU and one that uses just ALU. It should work in theory, but as you stated there would be the bandwidth problem. The only way, I can see a marked improvement would be using a dualie setup AND be able to assign a cpu. Opteron's with their on die cache would prolly be the best solution, the bandwidth concerns would be almost nil. HT is nice to have but in this case, I dont believe it would help alot other than using it on a single Project. (dont have a HT cpu, so no first hand knowledge, plz take with a grain of salt )

Force

 

[Ronin]

I can tell you that on my P4HT 3GHz, my avg WU time is about 1:45 (based on the information I garnerned from SetiSpy and the csv log). That's more than enough for me

 

[Unforgiven]

so 3:30 for 2 or is that just running 1 instance?

 

[Ronin]

3:30 for two is correct, yes, based on the #'s, give or take.

 

[kamper]

don't know about the fpu but there's more than one alu in a p4. so in theory if the processes could share the available alu's (which you couldn't program intentionally) it wouldn't matter much. but it seems to me you'd really have to dig into compiler design (assuming 2 projects are compiled with the same compiler) and intentionally optimize them to run with eachother on a ht machine. Good thought though Does anyone know if a ht processor actually has some dedicated circuitry per logical processor? Like are there circuits that would not be ever used if ht was disabled.

GunDog: dual opterons might be ideal but a single p4 with ht is much cheaper

 

[GunDog]

Originally posted by: kamper

GunDog: dual opterons might be ideal but a single p4 with ht is much cheaper

Yes, I agree with that. But the point I was going after was on ondie cache/fsb on the opterons.

 

[Ronin]

Actually, the dual Opterons that I have don't do Seti as well as I'd like. I'll check #'s again tomorrow, but they're a bit slower than some of the other crunchers I have (I have a Dual 240 setup).