With the current trends towards utilising cloud computing, it becomes
interesting to explore the utility of cloud computing in the area of high-
performance computing. Such area significantly differs from other typical
cloud applications in its intensive computation and communication
demands. This contradicts with the underlying sharing concept of
resources, exploited by cloud computing.
In this paper, we present work-in-progress for evaluating the performance
of the standard HPC SPECMPI benchmark on a private, dedicated, cloud
system. We compare the virtualised performance with native execution,
using the same underlying physical machines.
Generally, the cloud environment relies on virtualisation to allow for
consolidation of resources. This adds an overhead for emulating I/O
operations, as well as increasing the sharing overheads on the processors.
We construct a simple, yet efficient, analytical model that accounts for the
virtualisation overheads, taking into consideration the application
communication and computation patterns, and the underlying physical
computing and communication infrastructure. The developed model is a
queueing network model that accounts for the contention on physical
processing cores and the communication interconnect, and predicts
performance.
This model has the advantage of defining a basic model providing an
upper bound on performance of the SPECMPI on the cloud. Future work
will consider modelling other co-existing cloud applications. |
