[er_ijaz]

Dear all,


I have workstation 32GB, (4*8), 4 RAM sockets and each RAM is 8 GB
System monitor shows 8 CPU ( I think each core is divided in to 2 threads), whenever I try to run the simulation in multiprocessor,
It shows CPU is 100% but it would be using only around 8 GB of 32 GB memory,


I'm not sure why it is not completing utilizing the RAM.


Hence I removed 3 RAM sticks from the sockets, and only 1 RAM (8GB) is inserted, but I decompose to 16 processors and it runs with one RAM. Still system monitor shows 8 CPU and all CPUS are 100%, when I removed the RAM sticks still how it shows the CPU is 100% when there is no RAM itself.


I use Ubuntu 16.04

[flotus1]

Memory usage depends on the problem size. Not using all available memory is not an indicator for anything other than your simulation fitting into ram. Just as it should be. Removing memory will only make your system slower.
Imagine you have a mesh with 1 million cells and each cell needs to store 20 double precision floating point numbers. That's 1000000x20x8bytes of memory required. Using more than that usually does not make a code faster. Unless you chose different algorithms to speed things up which happen to require more memory.
CPU usage as reported by an operating system is not really a helpful indicator either. You can easily write a code that gets "100% utilization" on all cores but does not run any faster than on a single core.

[er_ijaz]

Hi


My concern is about number of processor,


Whether slow or fast is the next matter,



when I have only RAM (processor) how it can run in parallel....

[Simbelmynë]

I don't understand what you mean.


Please post information about which CPU you are using.


If you have a standard CPU, then you most likely have a dual channel setup, which means that you need to use two RAM modules to get the maximum solution speed.


With four physical cores on your CPU (hyperthreading turned OFF) then you should use a maximum of 4 cores in your parallel simulation. Anything more will slow down your simulation. In fact, 3 cores might be the optimum if you have other stuff going on while you run a simulation.


With only one memory module, you will most likely be memory-bandwidth limited as soon as you try a parallel simulation (single core might work without too much penalty, but probably not).

[flotus1]

Originally Posted by er_ijaz

when I have only RAM (processor) how it can run in parallel....

Not sure I understand the question. Could you rephrase it?
And from what I understood of your initial post, you are already running in parallel

[er_ijaz]

Hi I do apologies


I got confused on Number of RAM slots and Number of cores.


I was of thought number of RAM slots should be equivalent to number of cores. So each core will access particular RAM only.


So when a RAM stick is removed from its slot, that assigned core will not work,


I just understood it on repeated trial and errors

[LuckyTran]

Originally Posted by er_ijaz

Hi I do apologies


I got confused on Number of RAM slots and Number of cores.


I was of thought number of RAM slots should be equivalent to number of cores. So each core will access particular RAM only.


So when a RAM stick is removed from its slot, that assigned core will not work,


I just understood it on repeated trial and errors

No that's not how it works at all. Individual cores do not communicate with individual ram sticks. It should be quite obvious that there are 28-core CPU dies but there is no machine that has 28 ram sticks in it currently.

RAM is organized into banks and accessed via channels. Cores communicate with RAM through memory controllers. On one cpu-die, there is usually only one memory controller and all cores access the same RAM bank the same way. A RAM bank is like the street/road between CPU and RAM and the and the number of RAM channels is like the number of lanes on that street.

[er_ijaz]

Dear Lucky


Thank you for clarification