Avoiding congestion stopping prefetching

[frantgn90]

One important thing that SPP does is that in addition to throttling when Pd falls below the Tp, SPP also stops prefetching if there are not enough L2 read queue resources. Therefore, SPP does not issue prefetches when the number of empty L2 read queue entry becomes less than the number of L1 MSHR.

The first problem is that we have not access to L1 MSHR, so we have to try to apply another approach. Suggestions?

My problem is that I've not a clear idea about difference between L1 MSHR and L2 read queue because the first one holds the misses of L1 and these ones will go to the read buffer to L2, so when the movement is done?? Why we want two buffers??

[aztinki]

L1 cache misses and L2 read requests are the same

[aztinki]

So since we have L2 read queue we can use it

[aztinki]

We dont need two buffers

[aztinki]

I am sorry, I meant that both L1 misses and L2 read requests are inserted in the same queue for this Prefetching environment, I read the documentation. We will have to set a threshold ourselves, since we cannot see the L1 mshrs. Maybe tune it for what works best for our traces and spec traces

[frantgn90]

I set the threshold to 3/4 of the size of the L2 read requests buffer