Meaning of "l2_subp0_write_sysmem_sector_queries" event in nvprof

Hi,

I am profiling an application on GeForce GTX 1080. The application uses Unified Memory.

I was experimenting with the memory hints, and on profiling one of the kernels I am getting a large number (126203) for the event l2_subp0_write_sysmem_sector_queries. (I am assuming this is due to the hint “AccessedBy”) The description of this event says “Number of system memory write requests to slice 0 of L2 cache. This increments by 1 for each 32-byte access.

This statement is slightly confusing. Since L2 cache is on a lower level than system memory, the write request should be sent “from” L2 cache “to” system memory.

Similarly in the description for “l2_subp1_total_read_sector_queries” system memory is included as if it is on the same level as L1 and Texture cache. The definition says: “Total read requests to slice 1 of L2 cache. This includes requests from L1, Texture cache, system memory.

Can someone please clarify how this works? Or is there some issue in the way I am interpreting this definition?

Hi,

I have already raised your question to our developer.
I will reply as soon as I get the answer.

Thanks !

Hi veraj, thank you, looking forward to it.

System memory is cached in GPU L2 cache. So any system memory requests will go through the L2 cache. So the system memory write requests will be written to system memory via PCIE and nvlink. Also the read requests will be fetched over PCIE or nvlink if the system memory read requests misses L2. Hence the names are kept in that manner