5100 series compatible Motherboard.

[ivan_V_3]

I have a 5110P xeon phi coprocessor. I was wondering if the P9X79 WS motherboard will work with this. I emailed the company, and they seem to think it will, but it being a 300$ purchase, I would like a second opinion. 

I know P means passively cooled, but I am not worried about that part since I have plenty of fans, and it doesn't have to look pretty.

 

[ivan_V_3]

that works! I have a box fan on my pc, and it is taped up so that the air must escape through the card. Still, I am getting above critical temperature on my card. Ive tried the fan/duct system with a smaller fan, but it was not good enough. Do you know of any affordable fan/ducts for this system?

EDIT: nevermind that last question, I found your previous advice
https://software.intel.com/en-us/forums/intel-many-integrated-core/topic/537661

thank you so much for all your help. Im sure I will have questions in the future

[jimdempseyatthecove]

Woopie!

Now you have to work at getting the cooling under control.

The fan you use may have to be more capable than the one I used (you have an extra 75W to dissipate). See #8 for airflow requirements for 7120P. The fan has to meet the combination of cfm .AND. pressure drop. Most fan specs only list cfm.

The initial setup I used with the photos in the link above was not quite satisfactory under heavy load. Subsequently I replaced those fans with two higher capacity/higher pressure capable fans (higher RPM capable too). However, these were larger diameter and thus required a rework in the ducts. Sorry I do not have a picture of the ducts.

The ducts were fabricated out of vinyl signage (e.g. no parking sign) that I picked up at the local hardware store. I used vinyl (electrical) tape to hold the pieces together. I fabricated templates out of stiff paper which was easier to work with. Then cut the vinyl signage to match the templates.

Additional notes. See this.

The replacement ducts are (more or less) mounted on the case, one with inlet at front of case, one towards bottom, and are initially not connected to cards. After inserting cards, I then completed the connection with vinyl tape.

If you have a 3D printer, look at post #81 he may be willing to share is plot file.

The fan controller I first used could not supply sufficient number of watts per channel to get the (replacement) fans up to full speed. The initial fan controller, made by NZXT (don't have the number handy), form factor 5 1/4" full was replaced by an NZXT Sentry 2 5 1/4" half height. The former had more user control features and display but lesser watts per channel. The latter had more watts per channel and has been working well.

One of the links above references a different fan controller.

Whatever you come up with for a solution, please post it back here.

For others reading this thread.... Intel is now shipping its next generation Xeon Phi (Knights Landing). I expect that there will be a flood of the older generations Xeon Phi KNC cards on the secondary market at a relatively low cost. If you have a requirement of running applications suitable for use on Xeon Phi, then this might be a cost effective (but not necessarily headache free) to experiment with high-parallel/highly-vectorized coding. Then when you are ready, you can migrate to the KNL series with experience behind you and a basis for expectations.

Jim Dempsey

[ivan_V_3]

Jim,
The plan is to use 2x 
http://www.newegg.com/Product/Product.aspx?Item=N82E16811999612

(one push one pull) combined with some simple ducts made out of cardboard.

I tried my box fan Idea today, but It wouldnt go below 85 C. I have a feeling that it would work if it was winter and my house was at 15 C instead of 25 C. Either way, the fans should arrive in a week and I will post an update then. 

IV

[jimdempseyatthecove]

The Q&A section states 9.1 watts required to run at full speed (that is user statement). The NZXT Sentry 2 is spec'ed at 10 watts per fan channel. If your duct work is good, you should only need one of these fans.

Jim Dempsey

[ivan_V_3]

Alright so the 2 fans arrived, and you were right, the 7120P seems to hold at a steady 55 C with just one fan+duct system. I installed parallel studio XE so that I can run ICC but it is telling me ICC : command not found. When I check the supposed installation folder /opt/intel all I see is mic0. I have the export line in my .bashrc, but it doesnt mean anything if the folders arent there. I looked at the options in the installer but none tell me about the installation directory.

EDIT: solved the problem.

1. I needed to run the installer sh with sudo.

2. The path they suggested for export was fairly off. In my case it was

/opt/intel/compilers_and_libraries/linux/bin/intel64/

3. I also had to export:
export SINK_LD_LIBRARY_PATH=/opt/intel/compilers_and_libraries/linux/lib/mic:/opt/intel/compilers_and_libraries/linux/mkl/lib/mic

Also with my AC off, my card now hovers at around 60 C. This is with only 100 W being used.

IV

[jimdempseyatthecove]

Great. It looks like you are up and running.

60C is good. Try to construct a heavy load test that runs for 15 minutes or so. You want to keep the temperature under 83C. If it hits 85C the KNC CPU throttles back significantly. micsmc indicates when throttling occurs.

Now you have to learn how to use it, and as importantly, when to use it. When using offload programming model, the computation time of the offload has to significantly exceed the offload transfer/overhead + thread scheduling time (OpenMP, TBB, CilkPlus, ...) within the coprocessor. For native mode applications, you  may need to learn how to install the NFS drivers such that you can access the system or networked storage.

As a precautionary note, if you are using offload model, use thread pooling (such as OpenMP, TBB, CilkPlus, ...), if you are using pthreads, boost, std:thread, etc... then configure your threads to be encapsulated into a pooling system. IOW refrain from using join. The reason being is that the mpss software (in host and KNC) maintain per-process-per-thread-ID tables to manage packet routing. These tables are fixed size (on my system these were 512 entries per process). If you use spawn and join, each spawn generates a new thread ID, and you will quickly exceed the table size. This was experienced on an earlier MPSS. I cannot say if the newest MPSS vacates a table entry upon thread termination on host. If it does, then this is no longer an issue.

Jim Dempsey

[ivan_V_3]

Jim, 
I have been working on getting nfs working, but it is giving me issues.  

I created the folder on the host, and I forwarded it in /etc/exports. When I try to mount it on mic0, it tells me:

mount -o nolock -t nfs host:/home/vivan/hdr/shared /mnt/thehost
mount: RPC: Remote system error - No route to host
mount: mounting host:/home/vivan/hdr/shared on /mnt/thehost failed: Bad file descriptor

I've set up nfs on a cluster before, but this seems to be different.

IV

[jimdempseyatthecove]

Ivan,

I haven't setup the NFS. Use Google "intel xeon phi nfs mount", and start looking around. This one looks promising.

Did you install the NFS server on the Host?

Jim Dempsey

[jimdempseyatthecove]

Ivan,

If the above link does not get you going, then posting a new thread with the topic "How to perform NFS mount on KNC" you might get better responses.

However, first narrow the google search to the forum:

"intel xeon phi nfs mount site:software.intel.com"

That will give you Q&As of prior users experience

Jim Dempsey

[ivan_V_3]

Jim,
Thanks. It turned out to be a Centos nfs issue. I solved it by following the instructions at the bottom of

https://www.centos.org/forums/viewtopic.php?t=53896

IV