Is anyone out there running climateprediction on an Athlon X2 4800, 2500 MHz. with 2x512 k cache?

Having run hadcm3 on a Sempron 3400 for over a year I\'ve just upgraded to the faster Athlon X2 and two slab model instances (hadsm3fub) have arrived for me to run.

I\'m still using my old slow 533 MHz memory and possibly because of this the s/TS per core goes up significantly when I run two models, one on each core.

With one model on one core I get about 2.02 s/TS whereas with two on two cores I get about 2.32 s/TS on each, about 15% slower.

I wonder if anyone with the same processor running this model could let me know what memory they have what s/TS they are getting? This would enable me to estimate they effect of faster memory.

I\'m using Linux 2.6.22 but Windows figures would also help.

My processor is the G2 Brisbane core but figures for the earlier cores (larger cache, slower clock) would help too.

Many thanks,

Richard Driscoll

With the old version I was getting around 1.7 s/ts, but with the slower (and more careful) versions (5.40 and upwards) it is about 2.0 s/ts.

My (XP SP2) PC is overclocked to around 2560MHz, and the memory is DDR500 with T1 timings. Many people have noticed that running two (or four, on quads) climate models is slower than running one climate model and something else. This is because the climate models require considerable memory bandwidth.

---

I'm a volunteer and my views are my own.
News and Announcements and FAQ

A64 X2 4400+ with 2*1GB of Patriot DDR2 800 RAM, 2.0-3-2-5 1T (factory-certified SPD), running two Coupled Models:
2.35 s/TS v.5.44 CM
2.46 s/TS Spinup CM

---

"We have met the enemy and he is us." -- Pogo
Greetings from coastal Washington state, the scenic US Pacific Northwest.

With the old version I was getting around 1.7 s/ts, but with the slower (and more careful) versions (5.40 and upwards) it is about 2.0 s/ts.

My (XP SP2) PC is overclocked to around 2560MHz, and the memory is DDR500 with T1 timings. Many people have noticed that running two (or four, on quads) climate models is slower than running one climate model and something else. This is because the climate models require considerable memory bandwidth.

Riddle me this, batman. My computer here is a dual Opteron 248 HE at 2.2GHz with DDR400 ECC memory. How is it that I get better than 2.1 s/ts--faster than a Athlon X2 4800 @2.5GHz?

It\'s probably the internal differences in architecture of the cpus.

As far as I know the optimised HADCMs run rather faster on Linux than Windows. Could be another factor to take into account.

---

Cpdn news

Thanks everyone for the replies.

Mike
====

\"With the old version I was getting around 1.7 s/ts, but with the slower (and more careful) versions (5.40 and upwards) it is about 2.0 s/ts.\"

Sorry, will you explain this, I don\'t understand.


\"My (XP SP2) PC is overclocked to around 2560MHz, and the memory is DDR500 with T1 timings. Many people have noticed that running two (or four, on quads) climate models is slower than running one climate model and something else. This is because the climate models require considerable memory bandwidth.\"

Is your Athlon the pre-AM2 version with DDR memory rather than DDR2? Which one?

Are your 1.7 and 2.0 s/TS figs. with just one hadsm3 running?

Thanks,

Richard

---

As far as I know the optimised HADCMs run rather faster on Linux than Windows. Could be another factor to take into account.

Oh, I thought we were comparing HadSM models. If not, then I get about 2.2 s/ts on those HadCM models. That makes a bit more sense. I guess my advantage is that with two physical CPUs (vs cores on the Athlon), because the CPUs only have to share main memory (not cache). Could that be a reason for his reduced performance running two tasks?

Also, Richard is running Linux kernel 2.6.22, and I\'m running Linux kernel 2.6.21. It appears our setups are similar.

Thanks everyone for the replies.

Mike
====

\"With the old version I was getting around 1.7 s/ts, but with the slower (and more careful) versions (5.40 and upwards) it is about 2.0 s/ts.\"

Sorry, will you explain this, I don\'t understand.

There have been a number of versions of HadCM3. Prior to version 5.40 of HadCM3, the compiler was not inserting floating point checks after each calculation, but after this point the checks were added on the Windows builds. So on Windows, version 5.44 is slower than version 5.08 (but less likely to crash due to floating point errors from the PC). I don\'t think the Linux version performs these checks, but I may be wrong.

\"My (XP SP2) PC is overclocked to around 2560MHz, and the memory is DDR500 with T1 timings. Many people have noticed that running two (or four, on quads) climate models is slower than running one climate model and something else. This is because the climate models require considerable memory bandwidth.\"

Is your Athlon the pre-AM2 version with DDR memory rather than DDR2? Which one?

Are your 1.7 and 2.0 s/TS figs. with just one hadsm3 running?

These were DDR sticks rather than DDR2. As far as I can recall this was with two climate models running (I don\'t recall which particular one was the opposite). DDR1/500s sticks have roughly the same latency as DDR2/1000s.

I\'ve just noticed that your post was talking about HadSM3 not HadCM3 at the end.

On HadSM3 I get 1.56 s/ts on the PC with two running simultaneously.
http://climateapps2.oucs.ox.ac.uk/cpdnboinc/result.php?resultid=7031037

---

I'm a volunteer and my views are my own.
News and Announcements and FAQ

I had an Athlon 64X2 4200 (2,2 GHz) and 2 X 0,5 GB 667 MHz ram.
That machine used 2,25 and 2,29 s/Ts on hadcm-models.

I have AMD64 X2 6000+, 3.0GHz, 2GB DDR2 800 CL5, Windows XP Pro

With CM on one core and SM or AM on the other core I get:
HadSM3 1.29 s/TS = 588 credit/day
HadCM3 1.91 s/TS = 543 c/d
HadAM3 19.42 s/TS = 444 c/d