Re: [Gems-users] latency parameters in MOESI_SMP_directory protocol

[Niket Agarwal <niketa@xxxxxxxxxxxxx>]

So if all the network latencies are added separately from the directory latency, why the default parameter is quite high for DIRECTORY_LATENCY which is 80 cycles? I was thinking that's because of the location of the directory as you said which is assumed to be in memory.

Apart from going to off-chip to access memory to access the directory, there is also the latency to access the directory structure which contains list of sharers. Thus the 80 cycles. 

 

Also, other than parameters how do we really make sure that the directory is stored in memory or in the cache?

In SLICC, the directory look up happens in the _mem file. This ensures that directory structure is in memory.  

On the other hand, MEMORY_RESPONSE_LATENCY_MINUS_2 is set to 158 in the default parameters. Shouldn't that be set to something close to 80, since they are referring to same structure?

This is the latency to access the DRAM to get the actual data and is different than the directory structure. 

 

Accordingly, what would be realistic parameters for a 16-core CMP? Sorry for the long set of questions, again I appreciate any input for clarification.

This depends on your assumptions and configurations. You can look at related papers to what you are trying to model to get the best estimates. 

 

Thanks,

Ed


Also MEMORY_LATENCY is used to send the data to requestors.

On Wed, Aug 19, 2009 at 1:24 AM, Niket Agarwal <niketa@xxxxxxxxxxxxx> wrote:

I missed the 10 cycle link. Actual latency of access from core to directory = network delay + directory latency. In the above configuration, core0 to directory0 no contention latency = 1 cycle to get into Network interface + 1 cycle in Network Interface + 1 cycle in link to router + 4 cycle in router + 10 cycle in link to Network interface of directory + 1 cycle in network interface + 1 cycle to directory + directory latency.

On Tue, Aug 18, 2009 at 10:21 PM, Niket Agarwal <niketa@xxxxxxxxxxxxx> wrote:

 

Here the directory latency is set to 10 cycles and DIRECTORY_LATENCY parameter in rubyconfig.defaults is set to 80. So, what would be the actual directory latency observed here from core0 to directory0? Also, are directories stored in the main memory?

For SMP_directory protocols, the directory is stored in the memory by default. The latency of access from core to directory = network delay + directory latency. In the above configuration, core0 to directory0 no contention latency = 1 cycle to get into Network interface + 1 cycle in Network Interface + 1 cycle in link to router + 4 cycle in router + 1 cycle in link to Network interface of directory + 1 cycle in network interface + 1 cycle to directory + directory latency.
 

Thanks in advance for any insights,

Ed

_______________________________________________
Gems-users mailing list
Gems-users@xxxxxxxxxxx
https://lists.cs.wisc.edu/mailman/listinfo/gems-users
Use Google to search the GEMS Users mailing list by adding "site:https://lists.cs.wisc.edu/archive/gems-users/" to your search.

_______________________________________________
Gems-users mailing list
Gems-users@xxxxxxxxxxx
https://lists.cs.wisc.edu/mailman/listinfo/gems-users
Use Google to search the GEMS Users mailing list by adding "site:https://lists.cs.wisc.edu/archive/gems-users/" to your search.