[hpc-announce] Call for Submissions: Job Scheduling Strategies for Parallel Processing 2017

Walfredo Cirne walfredo at google.com
Thu Dec 22 11:06:38 CST 2016

Call for Submissions - jsspp.org
[image: JSSPP logo]

The JSSPP workshop addresses all scheduling aspects of parallel processing,
including cloud, grid (HPC) as well as “mixed/hybrid” or otherwise specific

Large parallel systems have been in production for more than 20 years,
creating the need of scheduling for such systems. Since 1995, JSSPP
provides a forum for the research and engineering community working in the
area. Initially, parallel systems were very static, with machines built in
fixed configurations, which would be wholesale replaced every few years.
Similarly, much of the workload was static as well, consisting of parallel
scientific jobs running on a fixed number of nodes. Systems were primarily
managed via batch queues. The user experience was far from interactive;
jobs could wait in queues for days or even weeks.

A little over 10 years ago, the emergence of large scale, interactive, web
applications together with the massive virtualization began to drive the
development of a new class of (cloud) systems and schedulers. These systems
would use virtual machines and/or containers to run "services", which would
essentially never terminate (unlike scientific jobs). This created systems
and schedulers with vastly different properties. Moreover, the enormous
demand for computing resources resulted in a commercial market of competing
providers. At the same time, the increasing demands for more power and
interactivity have driven scientific platforms in a similar direction,
causing the lines between these platforms to blur.

Nowadays, parallel processing is much more dynamic and connected. Many
workloads are interactive and make use of variable resources over time.
Complex parallel infrastructures can now be built on the fly, using
resources from different sources, provided with different prices and
quality of services. Capacity planning became more proactive, where
resources are acquired continuously, with the goal of staying ahead of
demand. The interaction model between job and resource manager is shifting
to one of negotiation, where they agree on resources, price, and quality of
service. Also, “hybrid” systems are often used, where the (virtualized)
infrastructure is hosting a mix of competing workloads/applications, each
having its own resource manager, that must be somehow co-scheduled. These
are just a few examples of the open issues facing our field.

>From its very beginning, JSSPP has strived to balance practice and theory
in its program. This combination provides a rich environment for technical
debate about scheduling approaches including both academic researchers as
well as participants from industry. JSSPP is a high-visibility workshop,
which has been ranking repeatedly in the top 10% of Citeseer's venue impact

Building on this tradition, starting this year, JSSPP also welcomes
descriptions of open problems in large scale scheduling. Lack of real-world
data substantially often hampers the ability of the research community to
engage with scheduling problems in a way that has real world impact. Our
goal in this new venue is to build a bridge between the production and
research worlds, in order to facilitate direct collaborations and impact.
Call for Papers

JSSPP solicits papers that address any of the challenges in parallel
scheduling, including:

   - Design and evaluation of new scheduling approaches.
   - Performance evaluation of scheduling approaches, including
   methodology, benchmarks, and metrics.
   - Workloads, including characterization, classification, and modeling.
   - Consideration of additional constraints in scheduling systems, like
   job priorities, price, accounting, load estimation, and quality of service
   - Impact of scheduling strategies on system utilization, application
   performance, user friendliness, cost efficiency, and energy efficiency.
   - Scaling and composition of very large scheduling systems.
   - Cloud provider issues: capacity planning, service level assurance,
   - Interaction between schedulers on different levels, like processor
   level as well as whole single- or even multi-owner systems
   - Interaction between applications/workloads, e.g., efficient batch job
   and container/VM co-scheduling within a single system, etc.
   - Experience reports from production systems or large scale compute

Call for Problems[image: NEW topic logo]

JSSPP also welcomes descriptions of open problems in large scale
scheduling. Effective scheduling approaches are predicated on three things:

   - A concise understanding of scheduling goals, and how they relate to
   one another.
   - Details of the workload (job arrival times, sizes, shareability,
   deadlines, etc.)
   - Details of the system being managed (size, break/fix lifecycle,
   allocation constraints)

Submissions must include concise description of the key metrics of the
system and how they are calculated, as well as anonymized data publication
of the system workload and production schedule. Detailed descriptions of
operational considerations (maintenance, failure patterns, fault domains)
are also important. Ideally, anonymized operational logs would also be
published, though we understand this might be more difficult. Scripts to
evaluate results and compute the metrics relevant for the system are highly

We envision that these papers will provide sufficiently detailed
information to be able to develop new scheduling approaches, which can be
robustly compared with the schedules used in production facilities, and
other approaches to solve the same problems.

For further information concerning paper formatting instructions please
visit the Submission section <http://jsspp.org/index.php?page=submission>.

Author submission deadline: February 1st, 2017
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.mcs.anl.gov/mailman/private/hpc-announce/attachments/20161222/282d68f6/attachment.html>

More information about the hpc-announce mailing list