Database Reference
In-Depth Information
- Higher task to worker ratios increase performance more for clusters that have
a small percentage of fast nodes than those with a small percentage of slow
nodes. In fact, for the largest data and a 75 % upgraded cluster, increasing the
task to worker ratio from two to four caused a 3.03 % execution time increase.
This is due to the overhead associated with the additional tasks. Whereas, for
a 25 % upgraded cluster a 10.36 % decrease in execution time was seen.
- The degree of heterogeneity is not only a factor of how many nodes are differ-
ent, but also the difference in computing ability of the various types of nodes.
This degree of heterogeneity can be used to help determine the optimial num-
ber of tasks that should be used to mitigate performance-heterogeneity in a
cluster.
The conclusions above are illustrated in Fig. 12 . This figure displays the aver-
age execution time per task, normalized based upon the data size. As we increase
the number of tasks, we can see that performance decreases when the cluster is
homogeneous due to the additional overhead associated with these tasks. Perfor-
mance improvements are not seen even though there are upgrades to the cluster,
as in the two and three tasks per worker cases, since the degree of heterogene-
ity physically provided by these upgrades is small. Despite the additional over-
head associated with generation of four tasks per node, we can see performance
improvements based upon the degree of heterogeneity within the cluster.
Fig. 12. This graph displays the results of our experiments as we increase the number
of tasks per worker and as we incrementally perform upgrades to subsets of the cluster.
The data here is normalized based upon data size, with the average execution time of
data per configuration presented on the Y-axis. On the X-axis is the number of tasks
assigned for the job in our 16 node cluster configuration.
 
Search WWH ::




Custom Search