[wrfems] Old DX servers

Douglas Tilly Douglas.Tilly at noaa.gov
Fri Dec 10 08:31:35 MST 2010


Hello all,

I finally had a chance to do some testing.  All I did was compare one of 
the DX servers to one of the LX machines.  These are not the really old 
LX machines, but those that we replaced just a month or two ago.  Here's 
what I found:

DX: 
x64 binaries - 1 hr 12 min
p7 binaries - 1 hr 15 min

So, even though it's supposed to be 64-bit chips on the DXs, using those 
binaries didn't seem to help. 

LX:
p7 binaries - 1 hr 6 min
core2 binaries - 1 hr 5 min

I haven't tried to use the core2 binaries on the DX yet, but I doubt 
there would be an improvement.  We're also guessing that the 800 MHz BUS 
is the bottleneck as that seems to be the biggest difference between 
these machines.  Needless to say, a very big disappointment. 

Doug


Tony Freeman wrote:
> Thanks Mark,
>
> Long story short: ... didn't help. Increased runtime a little (by 5 
> minutes but that's within my mental margin of error).
>
> I believe the biggest bottleneck in this whole setup is the 800 MHz 
> BUS speed.
>
> Anyway ... Long story long:  ... I checked the BIOS for each machine 
> and could not find a place to turn hyperthreading on/off ... so I 
> updated the BIOS on each machine from A01 to A07.  That didn't provide 
> a hyperthreading switch either but under BIOS --> CPU Info ... I could 
> enable/disable 'logical processors' ... so I disabled that.  Booting 
> up I found that my 4 processor machine became a 2 processor machine.  
> So I disabled 'logical processors' in all the BIOS's and made the 
> changes in /usr1/wrfems/EMS.cshrc and 
> /usr1/wrfems/util/benchmark/arw/conf/ems_run/run_ncpus.conf to account 
> for only 2 processors on each machine. 
>
>
> -- Tony
>
>
> Mark.Jarvis at noaa.gov wrote:
>> Tony:
>>
>> Your system information file shows Hyper-Threading turned ON.  That should be turned OFF.  The model does not run well with HT on.  See if you can turn that off in the BIOS and re-run the simulations.
>>
>> -Mark @ Louisville, KY
>>
>>
>> ----- Original Message -----
>> From: Tony Freeman <Tony.Freeman at noaa.gov>
>> Date: Wednesday, December 8, 2010 4:42 pm
>> Subject: Re: [wrfems] Old DX servers
>> To: WRF EMS Users <wrfems at comet.ucar.edu>
>>
>>
>>   
>>> I did a few runs today to establish a baseline. Here are the benchmarks:
>>>
>>>     * Running the benchmark on only the one machine I call 'clusterhead'
>>>       took 1 hour and 18 minutes.
>>>
>>>     * Running REALwrf on 'clusterhead' and ARWwrf on only the two nodes
>>>       took 36 minutes 35 seconds.
>>>
>>>     * Running REALwrf on 'clusterhead' and ARWwrf on all machines
>>>       (clusterhead, node1, node2) took 27 minutes 8 seconds.
>>>
>>>
>>> I'm guessing that if I had a fourth machine the runtime would be 
>>> quicker.  Attached is the benchmark output.
>>>
>>>
>>> The final run_ncpus.conf file has this setup:
>>>
>>> REAL_NODECPUS = clusterhead:4:192.168.0.1
>>> WRFM_NODECPUS = clusterhead:4:192.168.0.1,node1:4:192.168.0.43,node2:4:192.168.0.42
>>>
>>>
>>>
>>> -- Tony
> _______________________________________________
> wrfems mailing list
> wrfems at comet.ucar.edu
>   
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.comet.ucar.edu/pipermail/wrfems/attachments/20101210/b9d7084e/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: douglas_tilly.vcf
Type: text/x-vcard
Size: 345 bytes
Desc: not available
URL: <https://mailman.comet.ucar.edu/pipermail/wrfems/attachments/20101210/b9d7084e/attachment.vcf>


More information about the wrfems mailing list