Latency of remote dram controller

[shail...@gmail.com]

Hi Trevor,

I wanted to understand whether the latency of a dram request issued by a core and serviced from a remote dram controller same as that when serviced from the local dram controller?

For example, consider gainestown architecture with four cores A, B, C, and D and for dram controllers 1, 2, 3, 4 positioned at cores A, B, C, and D respectively. Now assuming that addresses are interleaved across dram controllers at cache line granularity, does Sniper model the latency of core A to controller 2 (remote) same as that to controller 1 (local)? I understand that Sniper has a simple memory model and returns a constant value as dram latency. However, is the variable network length from core to remote dram controller modelled?

Thanks,

Shailja

[Job]

Hi Shaijia,

I have the same problem as you, do you know the answer now, and if so, can you tell me?

[shail...@gmail.com]

Hi,

What I have understood is that all dram controllers are treated same in terms of latency. However, the different in the latency of the requests serviced from a local controller vs a remote one comes from the queue delay at the corresponding controller. 

Otherwise, all controllers (remote vs local) for a core return the same latency.

Thanks,

Shailja

[Job]

Hi Shaijia,

Then do you know how to set the queue delay of the different dram controllers(remote vs local)?

Thanks,

Job

[shail...@gmail.com]

You can take a look at DramPerfModelConstant::getAccessLatency() in dram_perf_model_constant.cc file. queue delay is computed using computeQueueDelay().

Thanks,

Shailja