ECE 6100 Project 3 : Implementation of Cache Coherence Protocols

Shantanu Kotkar December 6, 2013

Results
The simulator implemented in this project was used to run the given experiments. The results and statistics of the trace simulations for each of the cache protocol were recorded and analyzed to determine which protocol functions best in the given conditions. The experiment results are tabulated as follows:
RUN TIME MSI MESI MOSI MOESI MOESIF 219 317 217 217 217 $ MISS 8 7 7 7 7 EXPERIMENT 1 $ ACCESS SILENT UPGRADES 12 12 12 12 12 0 0 0 0 0 $-$ TRANSFERS 6 4 5 5 5

RUN TIME MSI MESI MOSI MOESI MOESIF 2367 2267 1167 975 683

$ MISS 30 30 30 31 34

EXPERIMENT 2 $ SILENT ACCESS UPGRADES 104 104 104 104 104 0 1 0 1 1

$-$ TRANSFERS 7 8 19 22 28

EXPERIMENT3 RUN TIME 3723 2607 3723 2607 1425 $ MISS 56 48 56 48 48 $ ACCESS 200 200 200 200 200 SILENT UPGRADES 0 8 0 8 8 $-$ TRANSFERS 20 23 20 23 35

MSI MESI MOSI MOESI MOSESIF

RUN TIME MSI MESI MOSI MOESI MOESIF 2265 1447 1869 851 551

$ MISS 27 19 29 19 19

EXPERIMENT 4 $ SILENT ACCESS UPGRADES 60 60 60 60 60 0 3 0 3 3

$-$ TRANSFERS 5 5 11 11 14

RUN TIME MSI MESI MOSI MOESI MOESIF 1661 1561 1261 1161 461

$ MISS 21 21 21 21 21

EXPERIMENT 5 $ SILENT ACCESS UPGRADES 37 37 37 37 37 0 0 0 0 0

$-$ TRANSFERS 5 6 9 10 17

RUN TIME MSI MESI MOSI MOESI MOESIF 7775 4925 6975 4125 3125

EXPERIMENT6 $ $ SILENT MISS ACCESS UPGRADES 87 62 87 62 62 747 747 747 747 747 0 25 0 25 25

$-$ TRANSFERS 12 15 20 23 33

RUN TIME MSI MESI MOSI MOESI MOESIF 6459 3993 5359 2909 2909

EXPERIMENT7 $ $ SILENT MISS ACCESS UPGRADES 79 55 79 55 55 952 952 952 952 952 0 24 0 24 24

$-$ TRANSFERS 17 17 28 28 28

RUN TIME MSI MESI MOSI MOESI MOESIF 9477 6441 8477 5241 4141

EXPERIMENT8 $ $ SILENT MISS ACCESS UPGRADES 110 92 110 92 92 800 800 800 800 800 0 19 0 19 19

$-$ TRANSFERS 18 30 28 42 53

Figure 1 shows the runtime (in clock cycles) of the trace simulations for each cache coherency protocol (MSI, MOSI, MESI, MOESI, MOESIF). Figures 2 and 3 show the number of cache misses and cache-to-cache transfers, respectively, that occurred during the simulation of each trace for each coherency protocol. Figure 4 shows, for each trace, the number of silent upgrades from the exclusive to the modified state.

Run time comparison (cycles)

10000 9000 8000 7000 6000 5000 4000 3000 2000 1000 0 MSI MESI MOSI MOESI MOESIF

Figure 1. Run-time comparison of the different cache coherence protocols for each experiment trace

Cache misses
120 100 80 60 40 20 0 MSI MESI MOSI MOESI MOESIF

Figure 2. Number of cache misses during simulation of the various traces using each cache coherency protocol.

Cache to cache transfers

60 50 40 30 20 10 0 MSI MESI MOSI MOESI MOESIF

Figure 3. Number of cache-to-cache transfers during simulation of the various traces using each cache coherency protocol.

Silent Upgrades
30 25 20 15 10 5 0 MESI MOESI MOESIF

Figure 4. Number of silent upgrades to the modified state from the exclusive state during simulation of the various traces using each cache coherency protocol that contains an exclusive state.

Observations
From the analysis of the comparison of run times of the traces for the cache coherency protocols, it is observed that the MOESIF protocol performs best, with the lowest runtime in each case, as can be seen from Figure 1. With the addition of the E state in the protocol, there is a steady decrease in the run-time. This fact can be attributed to the fact that as cache-to-cache transfers increase, the number of memory accesses decrease and this in turn reduces the latency to access the data from memory. However, it can be seen that experimental traces 2 and 5 do not conform to this trend. With respect to cache misses, the general trend is that it reduces with addition of the E state. However, once again experiment 2 does not follow this trend. Cache to cache transfers increase significantly in the MOESIF protocol, due to reduced number of memory accesses. From Figure 4. we know that the silent upgrades for MESI, MOESI and MOESIF remain the same throughout the experimentation.

Conclusions
Based on the results of the simulations, it is quite clear that MOESIF protocol performs the best among all the cache coherency protocols. Adding complexity to the protocol has in fact improved the performance of the protocol. The success of the MOESIF protocol is due to the

increased number of cache-to-cache transfers. This reduces the number of cache accesses and run time is reduced