68
0
10000
20000
30000
40000
50000
60000
70000
0
500
1000
1500
2000
2500
3000
Message Size (Bytes)
Message #
MG
Message Size Histogram
Figure 4.2:
MG Message Size Histogram
This figure shows the message size histogram for all 2615 messages sent by a single processor while
running the MG benchmark for class B input on 32 nodes. Message sizes are sorted in decreasing
order. Notice that message sizes follow an exponential distribution. Although most messages are
small, nearly all the data is sent in very large messages.
over 4 KB and 99% sent in messages over 1 KB. Although it has a larger number of small messages,
the total message cost is dominated by the large messages. The small message count is so low (around
1000) that these fail to have much impact on the total communication time.
4.2
Sensitivity Results
In this section, we examine the sensitivity of the NPB to the LogGP parameters. In partic-
ular, we concentrate the discussion on
and
. Given our MPI-GAM apparatus, and the nature of
the NPB sending a few long messages, these are likely to be the most important parameters. Recall
that for long messages, our apparatus inflates
for every 4KB fragment. Long messages will be af-
fected by
, although not nearly to the same extent as small messages. We will examine the validity
of this approach in the discussion section.
We attempt to explain any slowdown using the our knowledge of the MPI layer and the
simple piecewise-linear model of message passing in Section 2.3.2. Where we can not explain the
discrepancy, we hypothesize on reasons why a noticeable discrepancy exits between the measured