[Gems-users] irregular statistics due to the number of messages injected into the network by the coherence protocol

[Jesus Camacho Villanueva <jecavil@xxxxxxxxxx>]

Hi everybody!

We are simulating several applications in a sarek machine and using the
MESI_CMP_directory protocol.

At first, we simulated barnes and raytrace applications with different
topologies and we observed that the average (from 10 seeds) amount of
messages that were injected into the network differed in some cases more
than 50%!!! depending on the topology used.

Recently, we have run an FFT application with a basic network model that
introduces a constant message delay (that means we simulate the same
latency for each message going from any node to any other node) and we
have obtained the following results:

Constant latency (cycles) - execution time (cycles) - number of messages
- page_reclaims
1 - 3361858,8 - 830518 - 1018051,1
8 - 4346059,2 - 945225,1 - 1132667,2
16 - 5685143,7 - 958207,5 - 1145672,2
32 - 8410379,9 - 977663,1 - 1165158,3
64 - 14785582,6 - 1109518,3 - 1297078,7
128 - 20265856,9 - 594415,7 - 781911,1

10 different seeds have been used and average results plotted. As you
see, the execution time goes always up which is normal (the network
delay increases). We figure out the slight increase in the number of
messages and pages accessed is normal since the application is taking
more time. However, the severe drop when going from 64 cycles to 128
cycles was totally unexpected and we didn't find any reasonable
explanation, so we figure out there should be something wrong. The funny
thing is that the application finishes with half the pages claimed.

Anyone knows what it could be happening?

Any help will be really appreciated, since we are blocked with this issue!

Thanks a lot,

Jesús.