How well do these perform?

I curently have 4 computers ready to go, and well, 1 going now.

In 10 min, i have got .1% is that good ?

There overclocked to 3ghs, phenom 9950s on 750sb.

Hi Lukeg, welcome to the project and the forum.

We can\'t really comment on the speed of your models until they\'ve sent at least one trickle each and the trickles start to show up on their web pages. The best guide is the figure for seconds per timestep. You can expect the sec/ts ie the speed to be different for each type of CPDN model. As a guide, my Intel Core2Duo which is a fairly fast computer but not the fastest does between 14 and 19 sec/ts with HADAMs. Individual models can vary a bit in speed.

Here are the CPDN models I\'ve crunched on on the C2D. I haven\'t done many recently because I\'m working on Beta models, but I don\'t think the model speeds have changed much during the past year.

If you open the graphics for each model and press Z then 8 on the keyboard you\'ll hide the overlay and will be able to see more details including the speed as sec/ts. It\'s a cumulative average, not the current processing speed.

---

Cpdn news

What I recognised with my Intel Quad (and already with the Duo before) is the better speed if only one model is running at the time. Two or more models running simultaneous slow the WUs considerably, down to half the speed.

It would be interesting to know whether this happens with AMD as well.

---

Grüße vom Sänger

With a quad core phenom 9600 running Linux...
- running one hadsm3mh model and nothing else 0.72 s/TS
- running four hadsm3mh models ~1.05 s/TS each

What I recognised with my Intel Quad (and already with the Duo before) is the better speed if only one model is running at the time. Two or more models running simultaneous slow the WUs considerably, down to half the speed.

It would be interesting to know whether this happens with AMD as well.

I have an AMD Opteron 1210, which is a dual core, running six projects on Linux. Two projects can run simultaneously, and it may happen that CPDN and CPDN Beta run in parallel for a while. There is no slowdown.
Tullio

---

My (qualified) guess is that it boils down to memory bandwidth needed to run project work units.

If science app uses memory extensively (as it is in case of CPDN and some other memory-demanding projects), there are many cache misses. Which means that CPU needs to access main RAM. This is done through memory controller.

Here comes difference: Intel CPUs until very recently shared common memory controller (on north bridge) so adding up many cores posed a lot of stress to north bridge. It doesn\'t matter the number of physical CPUs vs. number of cores per CPU, it\'s the total number that matters.

Non-vintage AMD processors feature on-chip memory controller shared by all cores on same CPU. Adding up CPU chips actually increased overall RAM bandwidth as more memory controllers get added and with some help of OS\' process scheduler this meant smaller slowdown when running multiple science apps simultaneously[*]. Using many cores per CPU makes things similar to Intel case.

* There\'s still some slow down due to fact that probability that process is executed by one CPU while it\'s memory is physically allocated in RAM controlled by other CPU is non-negligible. That in turn means that there\'s overhead of communication between memory controllers, thus slowdown.

In short: in Intel case one gets same performance when using 2 dual-core CPUs as when using 1 quad-core CPU (given the same CPU frequency and the rest of HW). In AMD case one gets better performance when using 2 dual-core CPUs than when using 1 quad-core given proper OS support and proper RAM modules distribution.

Price tag is a completely different story though. SMP capable main boards and processors tend to be slightly more expensive than non-SMP capable counterparts.
One needs to read previous sentence thinking about number of chip packages and not about number of processors as seen by OS.

---

Metod ...