climateprediction.net home page
Contrubute to CPU stats=>help choosing CPUs

Contrubute to CPU stats=>help choosing CPUs

Message boards : Number crunching : Contrubute to CPU stats=>help choosing CPUs
Message board moderation

To post messages, you must log in.

AuthorMessage
bernard_ivo

Send message
Joined: 18 Jul 13
Posts: 438
Credit: 24,488,575
RAC: 2,962
Message 53560 - Posted: 2 Mar 2016, 16:04:14 UTC

Hi there,

Now and then one is trying to get a sense: "What CPU should I buy/use in my rig to increase performance/WU output of CPDN?" While there are many things to consider - base frequency, cores/threads, socket, price, energy use, floating point, CFP2006 rates.... a comparison between CPUs is kind of difficult.

My experience so far is that the most useful and somehow easiest comparison is to look at CPU's computational times for particular model (application).

Firstly I was looking at Average (sec/TS) per CPU type and CPDN model, but it was very time consuming and with great variances in results.

Later I found perhaps a better way provided by WUProp@home "a non-intensive project that uses Internet-connected computers to collect workunits properties of BOINC projects such as computation time, memory requirements, checkpointing interval or report limit."

Maybe if more CPDN crunchers join the project we can have better statistics on CPUs performance. Any thoughts, experience with WUProp or alternative views?
ID: 53560 · Report as offensive     Reply Quote
Les Bayliss
Volunteer moderator

Send message
Joined: 5 Sep 04
Posts: 7629
Credit: 24,240,330
RAC: 0
Message 53561 - Posted: 2 Mar 2016, 20:21:40 UTC - in response to Message 53560.  

cpdn is heavy on floating point maths. So:

It was, and perhaps still is considered that Intel chips are better than AMD chips for that.

Also, because of the various power saving features of laptops, a desktop is better.

The more recent the technology, and the more powerful the processor, the faster it should run climate models.
I have an Ivy Bridge and a Haswell machine, and the newer Haswell has been crunching models from the same batch type noticeably faster. So if/when I build a Skylake machine, it should be faster still.
(And the further up the specs in a given type, the faster too. e.g. I7 rather an an I3 or I5.)

Plenty of ram. At least 2 Gigs, and perhaps 3 Gigs per processor.
A sufficiently large psu, and a good brand of main board.

Finally, unless you have a very reliable power supply grid, and don't shut down often, (which can potentially result in loss of models), then there's a limit of the number of processor per computer that's efficient in terms of model success. The more processors, the more models that will/can/may get lost.
I don't know what this number is, perhaps around the 10-12 area?

ID: 53561 · Report as offensive     Reply Quote
Profile geophi
Volunteer moderator

Send message
Joined: 7 Aug 04
Posts: 2167
Credit: 64,403,322
RAC: 5,085
Message 53562 - Posted: 2 Mar 2016, 20:38:00 UTC - in response to Message 53561.  

Agree with Les. At this point, for both crunching speed and power efficiency, I would only consider Intel.

In a reasonably affordable range, a fast i7 with fast memory will get you the most throughput. I mainly crunch with AMD processors, but there is no doubt increasing the memory bandwidth and reducing the latency will improve throughput, especially for the more models that CPU runs. If you get something of the Haswell generation, DDR3 2400 is the sweet spot in terms of bandwidth per price. I know the Intel CPUs through Haswell "support" up to DDR3 1600, but that hardly is a limit, and better bandwidth does increase model throughput, especially as one increases core count.

If you're getting the newer generation of Intel, fast quad-channel DDR4 obviously helps.

Of course if you are thinking multi-socket hexa-core and beyond, you're out of my experience range and I don't know when adding additional cores reaches a level of significantly diminishing returns.
ID: 53562 · Report as offensive     Reply Quote
Profile Alan K

Send message
Joined: 22 Feb 06
Posts: 484
Credit: 29,579,234
RAC: 4,572
Message 53563 - Posted: 2 Mar 2016, 23:22:12 UTC - in response to Message 53562.  

I'd go a bit further on the RAM and say 2Gb per core. As an example my quad core i5 uses about 5.5Gb RAM when running 4 models.
ID: 53563 · Report as offensive     Reply Quote
Les Bayliss
Volunteer moderator

Send message
Joined: 5 Sep 04
Posts: 7629
Credit: 24,240,330
RAC: 0
Message 53564 - Posted: 2 Mar 2016, 23:29:56 UTC

Sorry, by "processor", I meant the number in each cpu chip.

Mine are 8: 4 real, 4 hyperthreaded.

ID: 53564 · Report as offensive     Reply Quote
bernard_ivo

Send message
Joined: 18 Jul 13
Posts: 438
Credit: 24,488,575
RAC: 2,962
Message 53566 - Posted: 3 Mar 2016, 8:36:49 UTC - in response to Message 53561.  

We do need to factor in the price as well. For example I would look into older and cheaper - around 450-500 USD (local prices) - set-ups like these:

Workstations
1) 2 x Intel Xeon Six-Core X5650 (2.66GHz, 12MB L3 Cache), (24GB) DDR3 ECC Registered (Up to 48GB), 250GB HDD Serial ATA / Integrated RAID 0, 1, 5, 10
2) 2 x Intel Xeon Quad-Core E5620 2.40GHz, 24576MB, 2 x 146GB, Dual LAN, 2 x PSU, 1U (650W PSU)
3) Intel Xeon Six-Core W3680 3.33GHz, 12288MB, 500GB

or rack servers
3) 2 x Intel Xeon Quad-Core X5550 2.66GHz, 32768MB, 4 x 73GB SAS, Dual LAN, P410i, 2 x PSU, 1U
4) 2 x Intel Xeon Quad-Core E5620 2.40GHz, 32768MB (24GB), 4 x 73GB, Quad LAN, 2 x PSU, 1U, (2 x 502W)

rather than i7 (Haswell or Skylake) as potentially the throughput of the above may be higher for lower prices - also comparing Average (sec/TS). Some of the old ones are around the limit of 10-12 cores (24 treads) as pointed by Les.

As for RAM I also think at least 2Gb per thread especially if the machine is used for other things while crunching. However it is still difficult to find user friendly stats that could easily facilitate a choice for CPDN machine.
ID: 53566 · Report as offensive     Reply Quote
Les Bayliss
Volunteer moderator

Send message
Joined: 5 Sep 04
Posts: 7629
Credit: 24,240,330
RAC: 0
Message 53567 - Posted: 3 Mar 2016, 8:55:54 UTC

Yeah, my figures were for the fastest, most reliable throughput of climate models.
So cost isn't a factor.

ID: 53567 · Report as offensive     Reply Quote

Message boards : Number crunching : Contrubute to CPU stats=>help choosing CPUs

©2024 climateprediction.net