src/mem/cache/cache.cc (Diff revision 1)

Is this correct?  Does not seem to agree with your comment that we take the maximum over all caches.

Show all issues

1. Andreas Hansson Sept. 22, 2015, 7:01 a.m. (Sept. 22, 2015, 7:01 a.m.)

   I'm pretty sure it's right.
   The max happens as part of the fanout in the crossbar.

src/mem/cache/cache.cc (Diff revision 1)

The trick here is in the difference between snoopDelay and headerDelay.
Snoop delay is where we do the max on each level. Header delay is just used to eventually pay for it and pass it "downwards".

Show all issues

1. Nilay Vaish Sept. 23, 2015, 9:41 a.m. (Sept. 23, 2015, 9:41 a.m.)

   Here is what I think you are doing. Suppose we have three levels of caching.  We have three private 
   
   controllers (A1,B1,C1) at the first level, another three private controllers at the second level 
   
   (A2,B2,C2) and one single shared controller (S) at the third level. When S receives a packet from 
   
   any of the second level controllers, it forwards  those packets to the other second level 
   
   controllers so that they can invalidate or writeback blocks if they need to.
   Now suppose A2 sent a packet to S which S forwarded to B2 and C2. Then, B2 and C2 would update the snoop delay
   
   of packet with their lookup latency and the header delay with the total delay of snoops sent to B1 and C1.
   
   Since all this was happening in parallel (which is why I think you take max), should not the header delay 
   
   be also maximum of the delays of snoops sent to B1 and C1.
   I might be completely incorrect here.

2. Andreas Hansson Sept. 23, 2015, 10:02 a.m. (Sept. 23, 2015, 10:02 a.m.)

   Who needs Sudokus... :-). I think you're right in that I've messed this up.
   The packet from A2 will be forwarded to B2 and C2 by the crossbar in front of S. First we forward it to B2, and B2 updates the snoopDelay with a max expression, and then takes a copy (snoopPkt), then forwards that to B1. On returning from the upwards snoop, the headerDelay from the copied snoopPkt is added to the original packets headerDelay (and not snoopDelay). I think this is indeed wrong, we should add it to the snoopDelay rather, and preferably do so before the max expression.

3. Andreas Hansson Sept. 23, 2015, 1:58 p.m. (Sept. 23, 2015, 1:58 p.m.)

   I broke out the upward snoop delay calculation, and handleSnoop now returns the delay. That way we can sum up the upwards delay (which in turn may be the result of a max), and the current cache, before taking the max and passing it downstream.

Ship It!