Throughout this section we will quantify disk virtualization using the term threads-per-disk (TPD). The total number of threads that can be executed on each cluster node
is the total number of hyperthreads the socket supports. For
the Atom cluster there are 4 threads/socket, and for the Core
i7 cluster, there are 8 threads per socket (4 cores/socket,
2 threads/core). The number of Hadoop Map and Reduce
slots that can be executed concurrently on one node of the
cluster is also a configurable parameter. All experiments on
our system are configured to run 2 Map and 2 Reduce tasks
simultaneously, utilizing all the 4 threads on a socket. The
Core i7 cluster cannot be configured to share disks among
cluster nodes. To approximate the effect of disk sharing
among multiple threads, we vary the number of Map and
Reduce tasks executing concurrently on the Core i7 nodes.
For example, with 4 Map and 4 Reduce tasks on one Core i7
node, the TPD ratio is 8. To achieve TPD of 4, the number
of Map and Reduce slots is configured to 2 each.