[QUOTE=YuL;388213]Thanks for the quick reply.
I plan to do it by running a 2x1 grid using MPI enabled msieve on a Dual Xeon E5-2620 @2.0GHz 32GB RAM,
can't wait to see how it will perform.[/QUOTE]

I think you will be disappointed.

[QUOTE=pinhodecarlos;388222]I think you will be disappointed.[/QUOTE]

Whyyyy ?
I'm currently running the post-processing of an SNFS job difficulty 234, I started LA 12 hours ago and the ETA is about 15h:
linear algebra completed 3808152 of 8581598 dimensions (44.4%, ETA 14h48m)

I think Pinhodecarlos is referring to the memory bandwidth, which is the limiting factor in the LA phase. With 12 cores across 2 sockets, the cores will most likely be starved for (memory) bandwidth.

For example:
My 2500k @4,0GHz scales almost linear from 1 to 3 cores, but the 4th core adds only single digit performance increase. This is largely due to the memory bandwidth bottleneck. For the record, that system has dual-channel DDR3-1600 memory (theoretically giving it 25.6GB/s).

Your system has a max memory bandwidth of 42.6 GB/s according to [URL="http://ark.intel.com/products/64594/Intel-Xeon-Processor-E5-2620-15M-Cache-2_00-GHz-7_20-GTs-Intel-QPI"]Intel Xeon E5-2620[/URL]. The extra L3 cache might also help a bit, but you'll also most likely be bandwidth limited.

[QUOTE=YuL;388224]Whyyyy ?[/QUOTE]
Not with why, but with what. With relative speed.
I.e. you may be disappointed that mpi on an on-board two-CPU system works similar to simply using more threads.

On multi-CPU systems (n>2), mpi seems to work better than plain multi-thread. I didn't see any advantage of mpi on a system just like yours (2x6 core Xeon).

I think it is safe to say that no one questioned that you will of course successfully finish algebra. It will be just at a normal, "no miracle" speed.

[QUOTE=Batalov;388251]It will be just at a normal, "no miracle" speed.[/QUOTE][YOUTUBE]ygE01sOhzz0[/YOUTUBE]

8.5M dimensions in 27 hours is a good speed but not astounding - it takes 30 hours on my single-socket i7/4930K. I was slightly hoping that the machine would be an E5-2620v3 with four channels of DDR4 per socket, which would have something like three times the theoretical memory bandwidth.

[QUOTE=fivemack;388260]I was slightly hoping that the machine would be an E5-2620v3 with four channels of DDR4 per socket, which would have something like three times the theoretical memory bandwidth.[/QUOTE]
I would like it, but I built that box 18 months ago, since an E5-2620v3 is $422, I'll wait a few months (years) before upgrading :)

[QUOTE=pinhodecarlos;388222]I think you will be disappointed.[/QUOTE]
For the moment I'm disappointed because I can't download the relations :grin:

Greg seems to be extremely busy at the moment.

[QUOTE=debrouxl;388350]Greg seems to be extremely busy at the moment.[/QUOTE]

[SPOILER]C'est bon il a répondu, je suis en train de télécharger les fichiers.[/SPOILER]

Just got his answer, the download is in progress.

Good :smile:

BTW, how do you want to be credited: "YuL", your real name, something else ?

[QUOTE=debrouxl;388375]BTW, how do you want to be credited: "YuL", your real name, something else ?[/QUOTE]

You can credit me with my real name: [SPOILER]Youcef Lemsafer[/SPOILER].