Hi Phil,
What network are you using? And how many nodes are there in your system?
The actual network delay can be much higher than the specified link
latencies.
Also, you might want to ensure that the RANDOMIZATION flag is set to
false. Otherwise messages get added "random" delays when enqueued into
the MessageBuffer. If you are running with Simics, the default is false.
-Niket
Philip Garcia wrote:
Hello,
I am trying to tweak my L2 cache miss latencies, however it seems that
whatever I do, I end up having a latency of over 40 cycles to my L2,
which is very unrealistic. I have set the standard latencys to be
relatively reasonable (and even tried ridiculously low values. Here
is a small trace of execution showing an issue to a memory address,
and the wakeup times (when enabling SLICC_DEBUG) that I see. I've
added a bit more to the outputs to show cycle times, and the latency
being added (for instance, the GETS request shows the latency it adds
before putting the requests on the request queue.
108079: L1 issuing GETS on chip 1 for addr [0xd388680, line 0xd388680]
new latency adding 2
108079: L1 issuing GETS on chip 1 for addr [0xd3886c0, line 0xd3886c0]
new latency adding 2
108081: ../protocols/MOESI_CMP_directory-L1cache.sm:331: MRM_DEBUG: L1
received
108081: ../protocols/MOESI_CMP_directory-L1cache.sm:332: WB_ACK_DATA
108081: ../protocols/MOESI_CMP_directory-L1cache.sm:331: MRM_DEBUG: L1
received
108081: ../protocols/MOESI_CMP_directory-L1cache.sm:332: WB_ACK_DATA
108099: ../protocols/MOESI_CMP_directory-L2cache.sm:885: [0xd388680,
line 0xd388680]
108100: ../protocols/MOESI_CMP_directory-L2cache.sm:885: [0xd3886c0,
line 0xd3886c0]
108120: ../protocols/MOESI_CMP_directory-L1cache.sm:654: MRM_DEBUG: L1
decrementNumberOfMessages
We recieved something for chip 1 on node 0 Addr:[0xd3886a0, line
0xd388680] Extra node handler working. It Took 43 cycles!
108122: ../protocols/MOESI_CMP_directory-L1cache.sm:654: MRM_DEBUG: L1
decrementNumberOfMessages
We recieved something for chip 1 on node 0 Addr:[0xd3886c0, line
0xd3886c0] Extra node handler working. It Took 45 cycles!
After running this, it appears that the L2 cache doesn't receive my
data for 20 cycles, and doesn't return the data back for another 20
cycles. I'm not sure which latency I'm missing here, but I thought an
L2 cache hit was just SEQUENCER_TO_CONTROLLER_LATENCY
+L1_REQUEST_LATENCY+L2_RESPONSE_LATENCY. I know there's also a little
added for network link latencies, but for now, I've set all of those
to 1.
If anyone knows something that I'm missing, I'd greatly appreciate
their help.
Phil
_______________________________________________
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.
|