Database Reference
In-Depth Information
Fig. 7.
Average number of matrices processed per second; averaged results across all
eight problem sizes. The X-axis displays performance heterogeneity in terms of the
percentage of the cluster that has been upgraded to Faster nodes. The Y-axis displays
thenumberof33
×
33 matrices multiplied per second in units of ten thousand. The
graph includes sets of bars for four different split granularities.
through all four bars in both cases. Likewise, it does not pay dividends for a
cluster with 75 % of its nodes upgraded. Within the 25 % and 50 % upgrade
levels, only the rightmost bar is taller, illustrating the need for enough (4) tasks
per node in the split to realize improved performance from the first 25 % of nodes
being upgraded.
Conclusions:
Cluster managers should not necessarily expect application per-
formance to improve at all due to partial upgrades, especially when the MapRe-
duce framework employs a traditional one-task-per-worker data split. Our results
in as displayed in Figs.
6
and
7
suggest that even MapReduce frameworks that
attempt to mitigate the effect of stragglers through the creation of additional
tasks may succeed only in adding overhead, and not decreasing runtimes when
they do not provide an adequate number of additional tasks. Such frameworks
can
, however, reap the benefits of partial upgrades with sucient split granular-
ities. In our tests, upgrading the first 25 % of nodes allowed MARLA to mitigate
the effect of stragglers well enough to have matrix multiply perform as well as a
more capable cluster that included 75 % Faster nodes.
6 Variability Between Upgrades
We introduce the following notation to facilitate discussion of this section's exper-
iments and results. A series of tuples,
<
p
i
,s
i
)
>
(
, describes the heterogeneity of
Search WWH ::
Custom Search