Speaker
Stefano Dal Pra
(INFN)
Description
Tier-1 sites providing computing power for HEP experiments are
usually tightly designed for high throughput performances. This
is pursued by reducing the variety of supported usecases and
tuning for performances those ones, the most important of which
have been that of single-core jobs. Moreover, the usual workload
is saturation: each available core in the farm is in use and
there are queued jobs waiting for their turn to run. Enabling
multicore jobs thus requires dedicating a number of hosts where
to run, and waiting for them to free the needed number of cores.
This drain-time introduces a loss of computing power driven by
the number of unusable empty cores.
As an increasing demand for multicore capable resources have emerged,
a Task Force have been constituted in WLCG, with the goal to define a
simple and efficient multicore resource provisioning model.
This paper details the work done at the INFN Tier1 to enable
multicore support for the LSF batch system, with the intent of
reducing to the minimum the average number of unused cores.
The adopted strategy has been that of dedicating to multicore a
dynamic set of nodes, whose dimension is mainly driven by the number
of pending multicore requests and fairshare priority of the submitting
user. The node status transition, from single to multi core et vice
versa, is driven by a finite state machine which is implemented in a
custom multicore director script, running in the cluster.
After describing and motivating both the implementation and the
details specific to the LSF batch system, results about performance
are reported. Factors having positive and negative impact on the
overall efficiency are discussed and solutions to reduce at most the
negative ones are proposed.
Author
Stefano Dal Pra
(INFN)
