1) Message boards : Number crunching : Boinc farms. (Message 15294)
Posted 1 Nov 2006 by ExtraTerrestrial Apes
Post:
Damn, they have cases! :o

MrS
2) Message boards : Number crunching : Boinc farms. (Message 15032)
Posted 9 Oct 2006 by ExtraTerrestrial Apes
Post:
>Maybe a new quad core AMD ...... ya ... thats the ticket.

I don't want to turn this in yet another AMD vs. Intel discussion, but if you want to buy before Q3 2007 (K8L) you should really consider a Kentsfield Core 2 Quad. In Q1 the cheaper one (600 - 700$?) with 2.4GHz is supposed to come out and should easily do 3GHz. This will smoke anything AMD can offer today in performance / Watt.

This being said, I'm still happily crunching with my 2.5GHz A64 X2 plus an AXP 2.16GHz as my *farm* :)

MrS
3) Message boards : Number crunching : Boinc farms. (Message 12193)
Posted 20 Jan 2006 by ExtraTerrestrial Apes
Post:
It's got a RAMBUS XDR mem controller integrated, so no FSB there. Connection to GPU is done over another very high bandwidth RAMBUS link (with some fancy name :D ). They knew they'd need an enourmous bandwidth to feed a beast like the Cell and this XDR gives them ... uhm, something around 50GB/s? I think it was said to be scalable up to 100GB/s, but they're not using the absolutely most expensive version. Oh, and I'm NOT talking about Mega-BITs.

MrS
4) Message boards : Number crunching : Boinc farms. (Message 12114)
Posted 17 Jan 2006 by ExtraTerrestrial Apes
Post:
Yep, relatively cheap yet very powerful. It's architecture would deserve its own thread, so just briefly: the current Cell has a simple dual thread PowerPC core for general tasks and feeding the "SPU"s. These 8 synergistic (or so) processing units are responsible for the high bandwidth number crunching. They're fairly simple (no multi threading here, 8 of them is enough already) in-order cores, specialized on vector operations.
In the PS3 they'll be using only 7 of them to improve yields. The tough job is to program that monster properly. I can imagine that in the worst case it could work on 7 seti-WUs seperately and should achieve an enormous throughput (for the price and it being only one single chip). Hope I my memory didn't fool me here ;)

MrS
5) Message boards : Number crunching : Boinc farms. (Message 12057)
Posted 15 Jan 2006 by ExtraTerrestrial Apes
Post:
I'm sure she will appreciate it if you let her do her rendering batch jobs and stuff on your farm ;)

MrS
6) Message boards : Number crunching : initial replication (Message 12053)
Posted 15 Jan 2006 by ExtraTerrestrial Apes
Post:
That's the way I like to see it. Now the question is how often that happens compared to our examples where all 5 were sent out immediately.

MrS
7) Message boards : Number crunching : Boinc farms. (Message 12049)
Posted 15 Jan 2006 by ExtraTerrestrial Apes
Post:
Amazing stuff guys! Especially all these old p3 500 - 1000MHz machines. Although I think I'd rather go with used parts for 130nm AXPs / Semporns and clock them somewhere over 2GHz. That should crunch at least the same amount of work and eat considerably less electricity. And then there's dual core Yonah nowadays, damn efficient but expensive because it's new.
I'm keeping an eye on the Playstation 3. It's supposed to be available with a hdd and linux on it, so even if the compilers are not top notch, it could just crunch 8 (7) different tasks and should have one hell of a throughput. That would temp me to get another box again :D
Oh, and when will someone develop a client for graphics cards?! The rumors for the new ATI look sweet: 48 pixel shader units, 32bit fp and relatively fast dynamic branching. Not that I want to start a huge discussion on that now, but that's what keeps me from building another system these days!

MrS
8) Message boards : Number crunching : initial replication (Message 12046)
Posted 15 Jan 2006 by ExtraTerrestrial Apes
Post:
If they are already adjusting parameters depending on the status of a study, then this (normal replication 3 or 4, later 5) should not be too hard to implement.

MrS

Edit & OT: finally added my avatar ... oh I love this picture :D
9) Message boards : Number crunching : initial replication (Message 12043)
Posted 15 Jan 2006 by ExtraTerrestrial Apes
Post:
@Michael:

That certainly makes sense. Why I'm asking is that in the beginning of a >500.000 WU job, the quick response is not needed and a lower quorum would increase the throughput (considerably). At the end of that job and for smaller ones the replication 5 is fine to get the latency of the results down.
I'd guess that this is just one parameter for the server which can be adjusted easily, maybe even automatically when they generate the jobs.

@Paul:

The way I understand it: initial replication means that they aim to issue 5 results. If there are not enough hosts demanding work, this may not happen, but normally it looks like this random one of mine:
http://lhcathome.cern.ch/workunit.php?wuid=988541

Like Desti already said, all 5 are sent out within 2 minutes. The quorum of 3 means that if they get 3 identical results and 2 differing ones, they will take the 3-result and won't send the WU out to other hosts.

MrS
10) Message boards : Number crunching : initial replication (Message 12021)
Posted 15 Jan 2006 by ExtraTerrestrial Apes
Post:
The search engine didn't find anything, so I assume this hasn't been discussed lately.

LHC uses an initial result replication of 5. IMO that's way too much, i.e. wasting a lot of CPU time. Einstein switched from 4 to 3 lately and seti is doing fine with 4.
What I suggest: use replication 3 in the beginning of a study, when there's still lot's of work to do and it doesn't matter if some WUs will take longer than 1 or 2 weeks. If the studies are almost finished, say 50.000 jobs left, you can go to replication 4 or 5 to finish the jobs quickly so you can start the final analysis.

I know LHC is doing fine with the CPU power they get, but all the time saved can possibly benefit other projects.

Regards, MrS



©2023 CERN