About the KNOB-"max_num_core_per_appl" and "max_threads_per_core"

[Applee]

Hi all,

For the parameter max_num_core_per_appl, my understanding is as the following:

     -- max_num_core_per_appl : the numbers of core to be assigned to one GPU application.  Is this right?

My params.in is like this: 

  max_num_core_per_appl = 1

  num_sim_small_cores = 4

  num_sim_medium_cores = 0

  num_sim_large_cores = 20

  max_threads_per_core = 80

Since I ran only one application, I only have one trace file as the following:

newptx                                                                                                                                

14

-1

/usr/local/src/gpuocelot/tests/cuda4.1sdk/.release_build/bd/ptx_jul2015/cuda4.1/QuasirandomGenerator/_Z26quasirandomGeneratorKernelPfjj_0/Trace.txt

----------------------------------

But the output is like this:

....

src/process_manager.cc:539: (I=0  C=0):  thread_count:1536

...

**Core 20 Thread 36 Finished:   insts:12268       cycles:1166718    (18667472)  seconds:76    -- 0.01 IPC (0.01 IPC) --  N/A  KHz      (0.16 KHz) 

...

**Core 21 Thread 34 Finished:   insts:11946       cycles:2302355    (36837664)  seconds:263   -- 0.01 IPC (0.01 IPC) --  N/A  KHz      (0.05 KHz)

...

**Core 22 Thread 10 Finished:   insts:11946       cycles:2302778    (36844432)  seconds:263   -- 0.01 IPC (0.01 IPC) --  N/A  KHz      (0.05 KHz)

...

**Core 23 Thread 46 Finished:   insts:11946       cycles:2303051    (36848800)  seconds:263   -- 0.01 IPC (0.01 IPC) --  N/A  KHz      (0.05 KHz)

...

**Core 21 Thread 372 Finished:   insts:12588       cycles:10453845   (167261504)  seconds:1590  -- 0.00 IPC (0.00 IPC) --  N/A  KH     z (0.01 KHz)

...

**Core 22 Core_Total  Finished:   insts:4330124     cycles:9848051    (157568800) seconds:1534  -- 0.44 IPC (0.44 IPC) --  N/A  KH     z (2.82 KHz)

...

**Core 23 Core_Total  Finished:   insts:4475270     cycles:10153031   (162448480) seconds:1571  -- 0.44 IPC (0.44 IPC) --  N/A  KH     z (2.85 KHz)

...

**Core 20 Core_Total  Finished:   insts:5051438     cycles:10155485   (162487744) seconds:1571  -- 0.50 IPC (0.50 IPC) --  N/A  KH     z (3.22 KHz)

...

**Core 21 Core_Total  Finished:   insts:4619456     cycles:10453845   (167261504) seconds:1591  -- 0.44 IPC (0.44 IPC) --  N/A  KH     z (2.90 KHz)

For this , I have some questions:

Sine I only ran one application, and assigned one GPU for it, why all the 4 GPU cores are running?  

Furthermore, the applications has 1536 threads, and the value of max_threads_per_core is set to 80. (1536 > 80 1536 > 80*4)  How can the number of threads exceeds the threshold?

Thanks for your time!

Best regards!

Applee

  

[Kim, Hyesoon]

First, max num core per appl  means how many applications per core  not how many cores per application. Sorry about this confusion. 

Second,  GPU has many CUDA blocks. 

In your setting, you have 4 + 20 cores, so you are using 24 cores.  Btw, please use only small cores for GPUS. Many code in macsim is hard-coded assuming that GPUs are small cores and CPUs  are using large  cores. 

Max thread per core specifies the number of threads per core that is running at the same time. The actual number of cuda threads (actually warps in CUDA applications)   running together  is set at run-time by CUDA applications (the number of allowable CUDA blocks per SM is recorded inside trace and the number of threads per cuda block is set by application) the max number is that it just cannot exceed that number, which is useful for CPU simulations. 

Core executes some number of cuda blocks and then if there are remaining cuda blocks, it will execute whenever a core has some space (I.e., when a block is finished) 

that’s why you see so many cuda threads have been executed. 

Hyesoon 

--

---
You received this message because you are subscribed to the Google Groups "Macsim Developer" group.
To unsubscribe from this group and stop receiving emails from it, send an email to macsim-dev+...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.