You are on page 1of 5

Vol. No. 1 Issue No.

1 International Journal of Interdisciplinary Engineering (IJIE) ISSN: 2456-5687

Compound Computing: A Simplified Scheduling


Mechanism for Enterprise HPC
Muda Rajesh Babu, Manager IT, VE Commercial Vehicles Ltd. New Delhi, India mrbabu@eicher.in

AbstractScheduling is complicated in terms of utilization of Grid computing advantage which becomes a reality
resources in High Performance Computing (HPC). In order to Get superior performance and computing
represent, it is recommended to divorce and identify the key
competences
areas of the problem. This paper mainly covers the approach for
providing correlation among one or more independent clusters. Use idle cluster to process the jobs
Using openlava as a lower-level scheduler, the complexity of the Use of multiple nodes to process a single job with
problem can be reduced. The major concern is to access MPI
uncertain resources and control the jobs exclusively for better Improve user job throughput
system utilization. Proposed solution showcase the requirement Add computing nodes in any cluster and utilize it
of the uncertain resources to be available in the enterprise from any location
middleware.
Strategy of computing resource implementation can
KeywordsCompound Computing; Scheduling; High be formed globally.
Performance Computing, Openlava; Workload Manager; Globus; Improvement in computational power without
Open Grid Service Architecture; Resource Management. affecting the budget.
I. INTRODUCTION Compound Computing enables a large organization to
form multiple collaborating clusters of computers such a way
HPC Clusters are the combination of more than one High that load sharing occurs not only within local clusters, but also
Performance Computing devices which are connected with in remote clusters.
each other via network to generate the output at extreme edge.
Compound Computing enables:
Openlava is a workload manager which schedules the user Load sharing among multiple clusters
jobs as per job submission policies and provide robust solution Co-scheduling across clusters: Job forwarding model
to manage the jobs in heavy workload condition. considers the statistics of remote cluster before
Compound Computing is a concept which is developed to sending jobs.
utilize more than one HPC clusters for equivalent prospect. Resource utilization and self-sufficiency of job
This concept will be feasible solution for the organizations management
which are having more than one HPC clusters around the Active Directory authentication mechanism and
globe. support for various file system
According to Sun model, idle time for cluster in U.S. will Straightforward file management over remote job
be the working time for India. So the Indian users can submit execution
their jobs to the U.S. cluster and vice versa. The advantage of
this concept is expansion of resource utilization. In case of III. ARCHITECTURE
unavailability of local cluster, user can submit their job to A. Flow of the Compound Computing:
remote cluster.
II. OVERVIEW OF COMPOUND COMPUTING Headnode 2

A. Purpose Compute 2.1

Within a single organization, sites may have distinct, Compute 2.2

autonomously managed openlava clusters. Having multiple Job Submission Headnode 1


(bsub)
openlava clusters could solve issues related to: Compute 1.1
Remote Cluster
1
Administration
Diverse geographic locations
Output Compute 1.2

Scalability User
Local Cluster
Headnode 3

B. Scope
Compute 3.1

When you have setup of more than one HPC clusters, it is


Compute 3.2
necessary to allow the clusters to cooperate to obtain the
following advantages of comprehensive load sharing ability: Remote Cluster

Access to a miscellaneous collection of computing 2

resources
Fig 1 Flow of Compound Computing

September 2016 Inside Journal (www.insidejournal.org) P a g e | 171


Vol. No. 1 Issue No. 1 International Journal of Interdisciplinary Engineering (IJIE) ISSN: 2456-5687
Fig. 1. Describes the flow of the compound computing A. Job spawning model
system. The job submission can be local or remote execution. In this model, the cluster that is looking for resources
If a job is submitted by default it initially reach the local sends jobs across the clusters. To work collectively, more than
cluster if found busy it tries to reach the other cluster that is one cluster must set up with openlava and compatible sender
idle. When a particular node is selected for the job submission and receiver queues.
then the job gets execution on that particular node, if resource
is busy it will wait in queue till requested resource is available. This model will help in scheduling remote jobs in three
After completing the job the files are copied back to the users scheduling phases: by default job will be executed on local
directory or desired location by mentioning the path. cluster; the local cluster opt a suitable remote receiver queue,
and send the job to it, then the remote cluster selects a
B. Architeture of Compound Computing appropriate node and forward the job to that node. The third
As per compound computing architecture, scheduling phase attempt to search remote cluster dynamically, the job
works based on local server and remote server. User can login will be submitted to any remote queue where resources are
on the local server using credential of active directory. Once available.
user will be authorized it will be redirected to Massive
B. Queuing System
Resource Broker (MRB) portal, which is built on top the
openlava scheduler. From the dashboard user can monitor the Sender queue: A sender queue will forward jobs to a
information of local and remote clusters. recognized remote queue. By default, Openlava tries to
execute job in the local cluster first. Openlava only tries to
schedule the job remotely if resources are not available in
Massive Resource
Broker Portal Compute Node 1 local cluster.
Resource Information
Compute Node 2 Storage Receiver queue: A receiver queue accepts the jobs from
Openlava Scheduler queues in a recognized remote cluster. Although sender
Compute Node 3
Head Node 1 queues only send jobs to appropriate queues in the remote
Local Server
Active cluster, receiver queues can accept the job from every queues
Directory in the remote cluster.
Massive Resource
Broker Portal Compute Node 1 Enable the Compound Computing Queues: For
Resource Information Storage
configuration of Compound Computing queues, do the
Compute Node 2
following:
Openlava Scheduler
Compute Node 3
Head Node 2 Send a job to any remote cluster: In the submission cluster,
Remote Server
configure a sender queue that forwards work to the all
available execution queue.
Begin Queue
QUEUE_NAME = send_remote
Fig 2 Architecture of Compound Computing
PRIORITY = 30
If resources are available in local cluster then user
NICE = 20
can submit their job locally. In second case, if the resources
are not available on local cluster then the job will be SEND_JOBS_TO = recieve@allclusters
transferred to the remote cluster. MRB will display the
HOSTS = none
information of local and remote clusters. Based on resource
availability user can submit their job on desired cluster. If user End Queue
is going to submit the job on remote cluster, then the inputs
file will be copied to the remote location and output will beA. In the execution cluster, configure a receiver queue that
generated at same location. If required, user can also specify accepts work from the cluster that contains the sender queue.
the output directory to local cluster or at his location. Begin Queue
IV. CONCEPT QUEUE_NAME = remote_accept
When a job is submitted to a cluster, it will be run locally SEND_JOBS_TO = queue2@cluster2 queue3@cluster3
or get execution on remote cluster. There will be a single
mechanism for resource sharing between clusters using RECEIVE_JOBS_FROM = cluster2 cluster3
Compound Computing. Here, Cluster 1 will submit the jobs to PRIORITY = 30
Cluster 2 or Cluster 3 using the job spawning model, and
queuing system will help user to forward the job to desire NICE = 20
cluster.
End Queue

September 2016 Inside Journal (www.insidejournal.org) P a g e | 172


Vol. No. 1 Issue No. 1 International Journal of Interdisciplinary Engineering (IJIE) ISSN: 2456-5687
Send a job to specific remote cluster: In the execution cluster, configure a receiver queue that
accepts work from the specific cluster.
In the submission cluster, configure a sender queue that
forwards work to the specific remote cluster. Begin Queue
Begin Queue QUEUE_NAME = queue2
QUEUE_NAME = queue1 RECEIVE_JOBS_FROM = cluster1
HOSTS = none DESCRIPTION = A receiver queue that receives the jobs
from cluster1.
SEND_JOBS_TO = queue2@cluster2
End Queue
MAX_RESCHED_TIME = infinite
DESCRIPTION = A remote queue that forward jobs to Queue1 in cluster1 sends all jobs to queue2 in cluster2.
cluster2. V. ALGORITHM
End Queue
Table 1 Algorithm

For (all queues)


{
If (number of pending jobs < MAX_RESOURCES && resource requirements satisfied)
{
If (SCHED_CONF == RESOURCE_ONLY)
{
Find queues with max (available slots)-(pending slots)
&& Forward job to this queue
}
Else
DEFAULT_PROJECTED set in SCHED_CONF
If ([(available slots) - (pending slots)] > 0)
{
If (SCHED_CONF == HIGH_PRIORITY_QUEUE)
{
Find queues with max queue priority
&&Find queues with max (available slots) - (pending slots)
&&Forward job to this queue
}
}
Else if ([(projected available slots) -- (pending slots)] > 0)
{
If (SCHED_CONF == PROJECTED_QUEUE_PRIORITY)
{
Find queues with min projected queue priority
}
Find queue with max (projected available slots)-(pending slots)
&&Forward job to this queue
}
Else if (SCHED_CONF == PENDING_WHEN_NOCPU)
{
Job pending
}
Else
{
Find queue with lowest (pending slots) / (total slots)
&&Forward job to this queue
}
}
Else
{
Job pending
}
}

September 2016 Inside Journal (www.insidejournal.org) P a g e | 173


Vol. No. 1 Issue No. 1 International Journal of Interdisciplinary Engineering (IJIE) ISSN: 2456-5687
PSEUDO Code: execution host, accordingly opt copy back the files to the user
Table 2 PSEUDO CODE location.
Step 1: Login
Step 2: Authentication through Active Directory
Step 3: Select the application from template
Step 4: Select the details for job execution environment like CPU,
execution host and other parameters required to execute the job
Step 5: Input file details
Step 6: Select type of execution from below list
i. Local Execution
ii. Remote Execution
iii. Dynamic Execution
Step 7: Select the files preference Fig 4 Selection of Execution
i. Copy back to users location
ii. Do not copy on users location
Step 8: Submit
Types of Execution:
VI. OUTCOME Fig.4 denotes the selection of execution modes.
After successful configuration of Compound
computing cluster, user can submit the job to local and remote Local Execution: the job will on local cluster and
cluster. wait till the required resources are available. The files will
Fig. 3. shows the job submission in MRB. store locally on the user location. This scenario is similar to
execute a job without a job.

Remote Execution: the job will execute on a remote


cluster though resources are available at local cluster. The files
will be copied back to user location on local cluster, if opted.
Users can thoroughly monitor their jobs on dashboard.

Dynamic Execution: the job will execute on local


cluster or remote cluster based on the resource availability.
The files will be copied back to the user location on local
cluster, if opted. Users can thoroughly monitor their jobs on
dashboard.

VII. CONCLUSION

In this paper, we have proposed concept of work


sharing among more than one HPC Cluster for better resource
utilization and improve the overall performance. Compound
Computing plays a crucial role in future to optimize the cost
of hardware and software resource at both institutional and
enterprise level.

Sun model gives flexibility to enable idle resources


Fig 3 Job submission through MRB
into employable. Optimum usage of license can be derived.
FLUENT gives us with a varied options like selecting The collaboration across the continents will be improved for a
the servers from the options that we have and giving a specific better future. Thus, organization can be benefited without
job name for a particular task and the selecting the version that hammering their budget using Compound Computing.
is compatible and user friendly. The queue works on the
option that is selected weather its a high priority job or ACKWNOLEDGMENT
normal job task that is based on the user discrimination.
Hostname avails us an option for selecting a particular desired Authors of this paper appreciate the provision of
node for running the job with the number of CPUs required. resources at VE Commercial Vehicles Ltd.
The journal and fluent files can be given as input expecting
the output in a desired file directory. Select the relevant

September 2016 Inside Journal (www.insidejournal.org) P a g e | 174


Vol. No. 1 Issue No. 1 International Journal of Interdisciplinary Engineering (IJIE) ISSN: 2456-5687
REFERENCE Efficiency in HPC on 2015 15th IEEE/ACM International
Symposium on Cluster, Cloud and Grid Computing
[1] S. Brunett,Center for Adv. Comput. Res., California Inst. of
Technol., Pasadena, CA, USA, K. Czajkowski ; S. Fitzgerald ; I. [4] K. I. Farkas ; Western Res. Lab., Digital Equipment Corp., Palo
Foster ; A. Johnson; C. Kesselman ; J. Leigh ; S. Tuecke Alto, CA, USA ; P. Chow ; N. P. Jouppi ; Z. Vranesic The
Application experiences with the Globus toolkit on High multicluster architecture: reducing cycle time through partitioning
Performance Distributed Computing, 1998. Proceedings. The on Microarchitecture, 1997. Proceedings., Thirtieth Annual
Seventh International Symposium on IEEE/ACM International Symposium on
[2] Deva Bodas, Justin Song, Murali Rajappa and Andy Hoffman Intel [5] H. Morohoshi ; Fac. of Comput. & Inf. Sci., Hosei Univ., Tokyo,
Corporation Simple Power-Aware Scheduler to limit power Japan ; Runhe Huang A User-friendly Platform for Developing
consumption by HPC system within a budget on 2014 Energy Grid Services over Globus Toolkit 3 on Parallel and Distributed
Efficient Supercomputing Workshop Systems, 2005. Proceedings. 11th International Conference on
(Volume:1 )
[3] Yiannis Georgiou, David Glesser, Krzysztof Rzadca and Denis
Trystram A Scheduler-Level Incentive Mechanism for Energy

September 2016 Inside Journal (www.insidejournal.org) P a g e | 175

You might also like