[Gems-users] Question about PT_TO_PT network latency.

[BYONG WU CHONG <bernard.chong@xxxxxxxx>]

Hello,

 

I have a question about GEM network latency.

Maybe this question is not really GEMS-related, but I would like some opinions from computer architecture people in GEMS mailing list.

 

I am using MESI_CMP_filter_directory for GEMS ruby simulation. I found out that this protocol, and maybe other protocols as well, uses PT_TO_PT network topology. And the network latency on that topology is chosen on default to be 14 cycles.

 

Is fixed setting of 14 cycles realistic to 2 core setting?

I thought it should be around 2 to 4 cycles.

 

If not, what will be the realistic network latencies for 2, 4, 8, 16, 32, 64 core setting?

 

I tried to get above numbers myself by using simple 1X2, 2X2, 2X4, 4X4 and 8X4 grid topologies with link_lat =2, router_lat = 3.

I know that this a wild assumption to figure out PT_TO_PT network latency from GRID network. If you can think of a better way to figure out the realistic PT_TO_PT NoC latency, please let me know.

I assumed sender and destination core is randomly decided with equal probability and this is the number I got.

 

Average network latency between cores

2 cores:   2 cycles

4 cores:   4 cycles

8 cores:   7 cycles

16 cores: 11 cycles

32 cores: 17 cycles

 

Do you think this is realistic enough to be used as PT_TO_PT NoC latency?

 

Thanks in advance.

 

- Bernard