[Ruby][CHI] send data two times for one single Read request in memory controller

[gxflying]

Hi sir:
in the implementation of CHI-mem.sm , I found that for a ReadNoSnp to ruby memory controller the data returned fron memory will be send two times when the controller state goes :
EVENT: --------ReadNoSnp--------------MemoryData-----------------------Trigger_Send-----------------------Trigger_SendDone
STATE : READY -----------> READ_MEM ------------> SENDING_NET_DATA -------------> SENDING_NET_DATA ------------------> READY
In both transitions triggered by MemoryData and Trigger_Send will ivoke sendDataAndCheck which will enqueue message into *m_datOut_ptr, and as a result , the msg will be send out from the datOut port twice ?!

((*m_datOut_ptr)).enqueue(out_msg, clockEdge(), cyclesToTicks(Cycles(m_data_latency)));

  transition(READING_MEM, MemoryData, SENDING_NET_DATA) {
    prepareSend;
    sendDataAndCheck;
    popMemoryQueue;
  }
  transition(SENDING_NET_DATA, Trigger_Send) {
    sendDataAndCheck;
    popTriggerQueue;
  }

gem5/src/mem/ruby/protocol/chi/CHI-mem.sm

Line 727 in d0d3c74

transition(READING_MEM, MemoryData, SENDING_NET_DATA) {

gem5/src/mem/ruby/protocol/chi/CHI-mem.sm

Line 733 in d0d3c74

transition(SENDING_NET_DATA, Trigger_Send) {

please help to figure out whether this behavior is expected ? thanks~

[powerjg]

@tiagormk do you know the answer to this question?

[tiagormk]

Yes, that's expected.
The number of messages sent depends on the amount of bytes being read an the size of the CHI data message. I believe the default size of the CHI data message is 32 bytes, and usually each request will read one cache line (64 bytes), so CHI-mem will split the data into two messages.

[gxflying]

Hi @tiagormk thanks for your clarificaiton
and I find the default size of mem channel size is indeed set to 32
what is the size is NOT 32 bytes? do it means we have to modified the source code in the aforementioned transition ？
for example 16 bytes channel size , we must modify the source code and call "sendDataAndCheck" 4 times (for a full cache line )?

gem5/src/python/gem5/components/cachehierarchies/chi/nodes/memory_controller.py

Line 69 in 461af51

self.data_channel_size = 32

[tiagormk]

it shouldn't require source code changes. sendDataAndCheck will be automatically triggered multiple times to send all necessary messages (see

gem5/src/mem/ruby/protocol/chi/CHI-mem.sm

Line 626 in d0d3c74

// end or send next chunk next cycle

).

[gxflying]

yeah, it seems that the sendDataAndCheck will handle multi beat data request,
then my question is what I asked the very fist,
why the flow will explicitly call the sendDataAndCheck twice？ while sendDataAndCheck can handle this automatically

1. one when transitoon from READING_MEM to SENDING_NET_DATA
2. one when SENDING_NET_DATA and triggered by Trigger_Send

  transition(READING_MEM, MemoryData, SENDING_NET_DATA) {
    prepareSend;
    sendDataAndCheck;
    popMemoryQueue;
  }
  transition(SENDING_NET_DATA, Trigger_Send) {
    sendDataAndCheck;
    popTriggerQueue;
  }

[gxflying]

Oh I see, the second sendDataAndCheck in my quesition is triggered by the sendDataAndCheck !
thanks I will close this issue~