Database Reference
In-Depth Information
the profile-based prediction. Furthermore, when faster nodes receive more work
and subsequently fail, the application stalls for longer than when slower nodes
fail.
Ahmad et al. [ 11 ] notice that Hadoop's speculative execution favors local
tasks over remote tasks, and speculatively execute remote tasks only near the
end of the map phase. The authors' Tarazu enhancement suite re-distributes
speculative execution of remote tasks based on an awareness of network commu-
nication. Tarazu monitors this communication to determine whether the map
phase or the shue phase creates the bottleneck. In contrast, our work delays
distributing data until just before workers need it, instead of making and then
reconsidering binding decisions. In addition, Tarazu considers clusters with two
classes of hardware, wimpy (Atom) nodes and brawny (Xeon) nodes. We find, as
shown by Nathuji et al. [ 14 ], that clusters more often have levels of hardware that
exhibit more closely-related performance than those considered for Tarazu. Our
work analyzes the scenario where the difference between worker nodes is a more
realistic depiction of the data center upgrade process described by Nathuji et al.
Our work studies how well a delayed task-to-worker binding of application
data to worker nodes allows a MapReduce framework to make ecient use of
performance-heterogeneous clusters, and the extent to which the strategy intro-
duces overhead in homogeneous clusters. We believe this paper to be unique in
varying both the granularity of data splits (and therefore the number of tasks),
and also the performance-heterogeneity of the underlying cluster.
3 Deferred Binding of Tasks
This paper characterizes the performance of delayed mapping of data and tasks
to worker nodes in the MARLA MapReduce framework. 1 This section describes
important MARLA features and distinguishes MARLA from Hadoop, primarily
with respect to how the two frameworks operate on performance-heterogeneous
clusters. We have described MARLA in more detail elsewhere [ 9 ], including its
performance improvements on load-imbalanced clusters.
Clusters whose nodes possess non-uniform processing capabilities (some nodes
faster than others) undermine Hadoop's strategy of partitioning data equally
across nodes and applying
methods uniformly. Workers on fast
nodes finish their work quickly but must wait for straggler workers on slower
nodes before the application completes.
MARLA works directly with existing cluster file systems instead of relying
on the Hadoop Distributed File System (HDFS) [ 12 ]. MARLA instead focuses
solely on
and
map
reduce
task management. MARLA uses a networked file sys-
tem (e.g. NFS) to decouple data management from the framework, allowing
the framework and the file system to address their separate concerns indepen-
dently. MARLA specifically targets high performance scientific compute clusters,
such as those at the National Energy Research Scientific Computing (NERSC)
and
map
reduce
1 MARLA stands for “MApReduce with adaptive Load balancing for heterogeneous
and Load imbalAnced clusters.”
Search WWH ::




Custom Search