Upload
ibm-india-smarter-computing
View
216
Download
0
Embed Size (px)
Citation preview
7/29/2019 How the IBM Platform LSF Architecture Accelerates Technical Computing
http://slidepdf.com/reader/full/how-the-ibm-platform-lsf-architecture-accelerates-technical-computing 1/11
1
Executive Summary Advances in High Performance Computing (HPC) have resulted in dramatic improvements inapplication processing performance across a wide range of disciplines that range from
manufacturing, finance, geological, life and earth sciences and many more. This mainstreaming of HPC has driven solution providers towards innovative Technical Computing solutions that are faster, scalable, reliable, and secure.
Today, these mission critical technical computing clusters are challenged with reducing cost and
managing complexity. Besides cost and complexity, data explosion in technical computing hastransformed compute-intensive application workloads to both compute and data-intensive. Therecontinues to be an unrelenting appetite to solve newer problems that are larger and even more
complex. This is straining technical computing environments beyond current limits. While today’s
technical computing application demands are growing, there are newer applications across several domains that now demand HPC scale solutions. These newer business problems include fraud detection, anti-terrorist analysis, social and biological network analysis, semantic analysis, drug
discovery and epidemiology, weather and climate modeling, oil exploration, and power grid management 1.
Although most technical computing environments are quite sophisticated, many IT organizationscannot fully utilize the available processing capacity in order to address newer business needs
adequately. For these organizations, effective resource management and job submission is anextremely complex process that needs to meet stringent service level agreement (SLA) requirementsacross multiple departments. This demands higher levels of shared infrastructure utilization and
better application processing throughput, while keeping costs lower. It is hard to optimize the
execution of a wide range of applications using clusters and ensure high resource utilization givendiverse workloads, business priorities and application resource needs.
To address these complex technical computing needs, IBM ® Platform™ LSF ® is successfully deployed
across many industries and is continuously evolving to address contemporary needs. The flagship product of the IBM Platform Computing portfolio, IBM Platform LSF provides comprehensive,intelligent, policy-driven scheduling features that enable users to fully utilize all their IT infrastructure resources while ensuring optimal application performance.
This whitepaper describes key architectural aspects of IBM Platform LSF including its use model, scheduling architecture, other core components and installation architecture. It highlights the
product’ s architectural strengths that help address current business challenges by optimizing the use
of shared HPC resources. The target audience includes chief technical officers (CTOs), technical evaluators and purchase decision makers, who need to understand the architectural capabilities of
LSF, and relate them to business benefits such as containing operational and infrastructure costswhile increasing scale, utilization, productivity and resource sharing in technical computing
environments.
1 Big Data in HPC – Back to the future http://blogs.amd.com/work/2011/04/13/big-data-in-hpc-back-to-the-future/
How the IBM Platform LSF Architecture Accelerates Technical Computing
Sponsored by IBM
Srini Chari, Ph.D., MBA
October, 2012
mailto:[email protected]
C a
b o t P a r t n e r s G r o u p , I n c . 1 0 0 W
o o d c r e s t L a n e , D a n b u r y C T 0 6 8 1 0
w w w . c a b o t p a r t n e r s . c o m
Cabot
PartnersOptimizing Business Value
7/29/2019 How the IBM Platform LSF Architecture Accelerates Technical Computing
http://slidepdf.com/reader/full/how-the-ibm-platform-lsf-architecture-accelerates-technical-computing 2/11
2
Introduction – Tuning Technical Computing Tasks
Advances in HPC and technical computing have resulted in dramatic improvements in application processing performance across a wide range of disciplines. Although most technical computingenvironments are quite sophisticated, many IT organizations find it challenging to maximize
productivity with available processing capacity and meet newer business needs adequately.
Today, HPC clusters typically consist of hundreds or thousands of compute servers, storage andnetwork interconnect components. These require substantial investment and drive up capital,
personnel and operating costs. For maximum Return on Investment (ROI), these technical computingenvironments must be shared across several users and departments within an organization. The ever
increasing computing demands in a continuously growing compute cluster requires fair sharing andeffective utilization of raw clustered compute capability. Sharing is made possible throughintelligent workload and resource management that includes job scheduling and fine grained control
over shared resources. Effective workload and resource management boosts cluster resourceutilization and Quality of Service (QoS) necessary for meeting business priorities and SLAs.
Technical compute cluster owners need to manage their existing deployed applications and also plan
for new business and application requirements. Maximizing throughput2 and maintaining optimalapplication performance are primary challenges that are hard to address simultaneously. Highthroughput requires elimination of load imbalance among constituent compute nodes in a cluster.Optimal application performance necessitates reduction in communication overhead byappropriately mapping application workload to the best available compute resources in the cluster.
Such needs are addressed by workload management solutions that typically consist of a resourcemanager and a job scheduler. Together, these prevent jobs from competing with each other for limited shared resources in large clusters.
IBM Platform LSF is a powerful and comprehensive technical computing workload management
platform that supports diverse workloads, across several industry verticals, on a computationallydistributed system. It has proven capabilities such as the ability to scale to thousands of nodes, built-in high availability, intelligent job scheduling and sophisticated yet simple-to-use resource
management capabilities that improve management of shared clusters. Features such as effectivemonitoring and fine-grained control over workload scheduling policies are well suited for multiple
lines of business users within an organization. By maximizing heterogeneous shared resources in ashared computing environment, LSF ensures that resource allocation is always aligned with business
priorities. System utilization and QoS improve as job throughput and application performance is
maximized. This reduces cycle times and maximizes productivity in mission critical HPCenvironments.
This whitepaper covers key aspects of the IBM Platform LSF architecture and how this architectureis optimized to address technical computing challenges. Highlights include key architectural aspects
of IBM Platform LSF including its use model, scheduling architecture, other core components andinstallation architecture that together help optimize the use of shared resources. This paper aims toempower CTOs, technical evaluators and purchase decision makers with a perspective on how the
architectural capabilities of LSF are well equipped to address today’s HPC challenges specific totheir business. Also included are the latest LSF features and benefits and how these help incontaining operational and infrastructure costs while increasing scale, utilization, productivity andresource sharing in technical computing organizations.
2 Throughput – number of jobs completed per unit of time
Technical
computing
environments
challenged tomaximize
productivity
Intelligent
workload and
resourcemanagement
are needed to
maximize ROI
and guarantee
stringent SLAs
IBM Platform
LSF
intelligently
schedules and
guarantees
completion of
workloads
across a
distributed,
heterogeneous,virtualized IT
environment
7/29/2019 How the IBM Platform LSF Architecture Accelerates Technical Computing
http://slidepdf.com/reader/full/how-the-ibm-platform-lsf-architecture-accelerates-technical-computing 3/11
3
The IBM Platform LSF Architecture
IBM Platform LSF provides resource-aware scheduling through its highly scalable and reliablearchitecture with built-in availability features. It has a comprehensive set of intelligent, policy-drivenscheduling capabilities that enable full utilization of distributed cluster compute resources. The LSFarchitecture is geared to address technical computing challenges faced by users as well as
administrators. Together with IBM Platform Application Center, LSF allows users to schedule
complex workloads through easy to use interfaces. With LSF, administrators can easily manageshared cluster resources up to petaflop-scale while increasing application throughput, maintaining
optimum performance levels, and QoS that is consistent with business requirements and priorities.Its modular architecture is unique and provides both higher scalability and flexibility by clearly
separating the key elements of job scheduling and resource management that are critical for HPCworkload management needs. These key elements are:
Task Placement Policies that govern exchange of load information within cluster nodes and are
used in decision making for task placement on cluster nodes
Mechanisms for transparent remote execution of scheduled jobs
Interfaces that support load sharing applications, and
Performance optimization of highly scalable HPC applications.
The following sections highlight the how LSF works, how users access its key features, the LSFscheduling architecture and its other core elements. Then, we briefly describe the installationarchitecture indicating where each LSF component is active within a cluster and how it helps in job
scheduling and resource management tasks.
LSF Cluster Use Model
This section describes how a typical IBM Platform LSF cluster is accessed and used. Individualcompute resources in a technical computing organization are usually grouped into one or more
clusters that are managed by LSF. Figure 1 shows this cluster use model, and how the jobmanagement and the resource management roles are played by different nodes in a LSF cluster. One
machine in the cluster is selected by LSF as the “master” node or master host. The master node playsa key role in resource management and job scheduling functions of workload management. Theother nodes in the cluster act as slave nodes and can be harnessed by the scheduler, through its
scheduling algorithms, for executing jobs.
Master Nodes: When nodes start up, LSF uses intelligent, fault-tolerant algorithms for master nodeselection. During system operation, if the master node fails, LSF ensures that another node takes the
place of the master, thus keeping the master node highly available and system services accessible tousers at all times. Job scheduling decisions are governed by business priorities and policies that areset up by the LSF system administrator.
Figure 1: LSF cluster use model (source: IBM)
Technical
computing
environments
challenged to
maximize
productivity
ntelligent
workload and
esource
management
are needed to
maximize ROI
and guarantee
tringent SLAs.
Platform LSF
ntelligently
chedules and
guarantees
completion of
workloads
across a
distributed,
heterogeneous,
virtualized IT
environment
he modular
BM Platform
SF architecture
rovides users
nd
dministratorsetter flexibility
nd scalability
ith separation
f scheduling
nd resource
anagement
ements
n intelligent
aster-slave
odel for
cheduling and
anagement
mproves
eliability and
erformance
7/29/2019 How the IBM Platform LSF Architecture Accelerates Technical Computing
http://slidepdf.com/reader/full/how-the-ibm-platform-lsf-architecture-accelerates-technical-computing 4/11
4
Users connect to a distributed system via a client and submit their jobs to the job submission node.As these user jobs queue up, the master decides where to dispatch the job for execution, based on the
resources required and current availability of the resources among slave nodes.
Slave Nodes: Each slave machine in the system collects its own “vital signs” or the load information periodically and reports them back to the master. Detailed information on the load index3 for eachnode in the distributed system is analyzed and used for scheduling decisions in order to reduce job
turnaround time and increase system throughput. LSF has unique algorithms for smart informationdissemination of the load index and resource usage status to optimize system scalability and
reliability. These algorithms are proven to scale up to thousands of nodes.
Workload Execution: LSF has a remote execution component that starts or stops the jobs on the
assigned slave node. Once the scheduled jobs complete on slave nodes, the completion results and job status are communicated to the user. LSF also generates reports on resource usage and detailed
job execution logs. Users can obtain job execution results on a local node, transparently, as if theywere executing those jobs locally. LSF frees users from having to decide which nodes are best for
executing a job while allowing administrators to set up policies for job execution logic that are bestsuited to business needs.
There are options to checkpoint a job that is running on a slave node, or move a running job to adifferent slave node and then resume execution. This feature can help to temporarily suspend
running jobs, free up resources for any critical jobs, and then resume jobs from the last execution point instead of having to restart them all over, thus improving system flexibility and utilization.
LSF Scheduler
Scheduling is a key component of any workload and resource management solution. Figure 2 shows
the central component of the LSF scheduling architecture, which provides support for multiple
scheduling policies. When a job is submitted to LSF, many factors control when and where the jobstarts to run. These factors include the active time window of the queues or hosts, resourcerequirements of the job, availability of eligible hosts, various job slot limits, job dependencyconditions, fair-share constraints and load conditions.
3 Load Index: LSF defines a load -index for each type of resource. Load index quantifies each node’s loading condition. Depending on the nature of the
resource, some possibilities are queue length, utilization, or the amount of free resource. Reference: Utopia – a load sharing facility for a large scale
heterogeneous system
http://cse.unl.edu/~lwang/project/Utopia_A%20Load%20Sharing%20Facility%20for%20Large,%20Heterogeneous%20Distributed%20Computer%20Syst
ems.pdf
Figure 2: LSF scheduling architecture (source: IBM)
Smart
scheduling
algorithms
reduce time to
results and maximize
throughput
while
improving
reliability
The LSF
scheduler
supports
multiple
policies
aligned with
business needs
7/29/2019 How the IBM Platform LSF Architecture Accelerates Technical Computing
http://slidepdf.com/reader/full/how-the-ibm-platform-lsf-architecture-accelerates-technical-computing 5/11
5
One unique architectural feature of the LSF scheduler is that it allows multiple scheduling policies tocoexist in the same system. This means that to make scheduling decisions, LSF accommodates
multiple scheduling approaches that can run concurrently and be used in any combination, includinguser-defined custom scheduling approaches. The LSF scheduler plug-in API can be used to
customize existing scheduling policies or implement new ones that can operate with existing LSFscheduler plug-in modules. These custom scheduling policies can influence, modify, or override LSFscheduling decisions, thus empowering administrators to model the job scheduling decisions aligned
with business priorities. The scheduler plug-in architecture is fully external and modular; newscheduling policies can be prototyped and deployed without changing the compiled code of LSF.
LSF Core Components
LSF takes job requirements as inputs, finds the best resources to run the job, schedules and executes jobs and monitors its progress. Jobs always run according to host load and site policies. This section
provides an overview of some of the core components of LSF and their key role in job schedulingand resource management functions. LSF is a layer of software services on top of UNIX and
Windows operating systems that creates a single pool of networked compute and storage resources.
This layered service model (Figure 3) provides a resource management framework to allocate,manage and use resources as a single entity. The three basic components of this layer are LSF Base,LSF Batch and LSF Libraries and together they help in distributing work across existing
heterogeneous IT resources; creating a shared, scalable, and fault-tolerant infrastructure that deliversfaster and more reliable workload performance.
LSF Base provides basic load-sharing services for the distributed system such as resource usageinformation, host selection, job placement advice, transparent remote execution of jobs and remote
file options. These services are provided through the following sub-components: Load Information Manager (LIM)
Process Information Manager (PIM)
Remote Execution Server (RES)
LSF Base application programming interface (API)
Utilities such as lstools, lstcsh and lsmake.
LSF Batch extends LSF base services to provide a batch job processing system along with load balancing and policy-driven resource allocation control. To provide this functionality, LSF Batch
uses the following LSF base services:
Figure 3: LSF services - high level architecture (source: IBM)
The LSF
scheduler
minimizeslatencies for
short jobs
while
improving
performance
for long jobs
LSF core
components
help in
distributing
work across
existing
heterogeneous
IT resources;creating a
shared,
scalable, and
fault-tolerant
infrastructure
7/29/2019 How the IBM Platform LSF Architecture Accelerates Technical Computing
http://slidepdf.com/reader/full/how-the-ibm-platform-lsf-architecture-accelerates-technical-computing 6/11
6
Resource and load information from LIM to perform load balancing activities
Cluster configuration information and master LIM election service from LIM
RES for interactive batch job execution
Remote file operation service provided by RES for file transfer.
LSF Libraries provide APIs for distributed computing application developers to access job
scheduling and resource management functions. There are two LSF libraries: LSLIB and LSBLIB.
LSLIB is the core library that provides basic workload management services to applicationsacross a shared cluster and is a runtime library to easily develop load sharing applications.
LSLIB implements a high level procedural interface that allows applications to interact withLIM and RES. The other library, LSBLIB, is the batch library and it provides batch servicesthat are required to submit, control, manipulate, and queue jobs on system nodes.
LSF Installation Architecture
LSF consists of a number of servers or daemon processes that run with root privileges on each participating host (Figure 4) in the system and a comprehensive set of utilities that are built on top of the LSF API. There are multiple LSF processes running on each host in the distributed system. The
type and number of processes running depend on whether the host is a master host, a compute or slave host or one of the master node candidates as shown in Figure 5.
LSF libraries
provide APIs
for applicationdevelopers to
access job
scheduling and
resource
management
functionality of
LSF.
LSF consists of
a number of servers or
daemon
processes that
run with root
privileges on
each
participating
host
Figure 4: LSF daemons and their functions in scheduling & resource management (source: IBM)
7/29/2019 How the IBM Platform LSF Architecture Accelerates Technical Computing
http://slidepdf.com/reader/full/how-the-ibm-platform-lsf-architecture-accelerates-technical-computing 7/11
7
On each participating host in a LSF cluster, an instance of LIM runs and exchanges load informationwith its peers on other hosts and provides applications and associated tasks with a list of hosts that
are best for execution. Multiple resources on each host and resource demands of each application areconsidered in LIM placement decisions. In addition to help LSF make placement decisions, LIM
also provides load information to those applications that make their own placement decisions.Besides LIM, RES is another server or daemon on each host. RES provides the mechanisms for transparent remote execution of arbitrary tasks. Typically, after placement advice has been obtainedfrom LIM, a stream connection is established between the local application and its remote task
through RES on the target host. This is followed by remote task initiation. LSF supports several
models of remote execution to meet the diverse functional and performance requirements of applications. A LIM and a RES run on every Platform LSF server host. They interface with thehost’s operating system to give users a uniform, host-independent environment. Figure 6 shows
sample job submission steps, for regular as well as batch jobs that run on a LSF system and variousinteractions between LSF components during job submission and execution.
LSF Architectural Strengths
The architectural strength of LSF results from its modular structure that even allows parts of thesystem to be used independent of other parts. For instance, a task can be executed on a remote host
specified by the user so that LSLIB can contact the remote RES component, without needing theLIM component. Similarly, load information and placement advice from LIM may be obtained for
Figure 5: Installation architecture with various LSF processes running on different nodes in a LSF managed cluster (source:
Figure 6: Interactions between various LSF components during job submission and execution (source: IBM)
LSF supports
several models
of remoteexecution to
meet the
diverse
functional and
performance
requirements
of applications
The LSF modular
structure even
allows parts of
the system to
be used
independent of
other parts
7/29/2019 How the IBM Platform LSF Architecture Accelerates Technical Computing
http://slidepdf.com/reader/full/how-the-ibm-platform-lsf-architecture-accelerates-technical-computing 8/11
8
purposes other than remote execution. Another advantage of the LSF architecture is that policies andmechanisms of load sharing may be changed independent of each other as well as independent of the
applications running on the system. This provides significant fine grain control over resource sharingand job scheduling.
While LSF manages distributed system sharing and job scheduling complexities with its smartarchitecture, it also provides easy-to-use and simple interfaces that improve productivity of both
users and administrators and boosts collaboration in technical computing organizations. The highlyavailable single master node concept for managing an entire cluster simplifies distributed systems
management and frees up domain experts to focus on value added work instead of the tedious jobscheduling and system management tasks. At higher scale, LSF deploys a hierarchical master nodeconcept internally but all that complexity is hidden and does not impact its simplified use model.
Users can access systems with thousands of nodes that could be spread across geographies throughadditional LSF components such as LSF Multi-Cluster. LSF is architected to run on a variety of x86
hardware and operating environments including the latest generation of IBM System x servers and isalso certified on IBM Power Systems servers running the AIX and Linux operating systems.
IBM Platform LSF Benefits
LSF allows multiple users to share heterogeneous assets more effectively in a shared computingenvironment.
Consequently, people are more
productive, projects are completedearlier and because computer utilization is better, infrastructure
costs are contained.
By consolidating compute resourcesfrom multiple, distributed systems,workload can be distributed more
efficiently across an organization’stechnical computing assets that aregeographically dispersed. With thiscapability, effective sharing of resources can be extended from a
single cluster to enable flexiblehierarchical or peer-to-peer workloaddistributions between multiple clusters.
LSF improves efficiency by removing the problem of underutilized compute resources by enabling
local administrators to retain control of their own assets while still permitting remote systems to tapinto idle capacity.
Cluster-level capabilities in LSF transparently extend to the grid. This makes it exceptionally fast andcost-efficient to deploy on grids, eliminating the need for sites to implement an expensive,customized scheduling layer to share resources between clusters.
With simple interfaces and a plug-in modular architecture, LSF lowers the learning curve andincreases cluster user productivity, reduces application integration and training costs, and speeds up
job completion by eliminating manual job submission errors through automation. Technicalcomputing users obtain faster results and complete more jobs using shared cluster resources at lower costs.
BM PlatformLSF: Complete,powerful,scalable
Workload Management Solution.
Benefits:Advanced,
feature-rich
workloadscheduling
Robust set of add-on
features
Integrated
applicationsupport
Policy &
Resource
awarescheduling
Resource
consolidation
for maximumperformance
Automation &
Advanced self
management
Thousands of
concurrent
users & jobs
Optimalutilization, less
infrastructure
costs
Better user
productivity,
faster time to
results
Best TCO –
flexible control,multiple
policies, robustcapabilities,
administratorproductivity
The LSF
smarter
scheduling
advantages:
Higher throughput at alower cost
Flexibility toaddresschanging
business needs
Better assetutilization &ROI
Better servicelevels to end-users
Increasedautomation &
reducedmanualintervention
7/29/2019 How the IBM Platform LSF Architecture Accelerates Technical Computing
http://slidepdf.com/reader/full/how-the-ibm-platform-lsf-architecture-accelerates-technical-computing 9/11
9
In short, LSF equips technical computing environments to achieve the following benefits:
Obtain higher-quality results faster
Reduce infrastructure and management costs, and
Easily adapt to changing user requirements.
Conclusions
Flexibility, scalability and agility are the key requirements of technical computing environments4.Technical computing users typically run varied applications and workloads on clusters and largescale distributed systems. These workloads range from performance sensitive, compute-intensive,data-intensive or a combination.
To support large technical computing clusters, customers are challenged with manual tasks andcumbersome tools, issues related to integration and the need for multiple dedicated personnel todevelop and maintain custom integration between various tools and applications. This increases costs
and business risks because a lot of the mission-critical functionality could be expensive or time-consuming to realize. Instead of focusing on core high-value tasks, administrators could also be
consumed by mundane manual systems management tasks. These environments demand reliability aswell as scalability from the underlying IT infrastructure. However, budgetary constraints and
competitive pressures make it imperative to increase resource utilization and improve infrastructuresharing efficiencies to achieve better collaboration, productivity and faster time to results.
In such large scale distributed systems, computing resources are made available to users throughdynamic and transparent load sharing provided by IBM Platform LSF. Through its transparent
remote job execution, LSF harnesses powerful remote hosts to improve application performance,enabling users to access resources from anywhere in the system . The IBM Platform LSF productfamily has the broadest set of capabilities in the industry which are tightly integrated and fullysupported by IBM. As part of an even broader portfolio of offerings from IBM and IBM BusinessPartners, LSF can be packaged with more engineering, integration and process capabilities. This
further enhances productivity of technical computing users, enabling them to focus more on their core business, engineering or scientific tasks. It also reduces future strategic risk as the business
evolves.
The IBM Platform LSF architecture is geared to create a scalable, reliable, highly utilized and
manageable shared infrastructure for technical computing environments with powerful resourcemanagement and scheduling solutions cutting across cluster silos. Its modular architecture provides
the much needed flexibility and fine-grained control while speeding up job turnaround times andimproving productivity. Simple interfaces and easy customization features of LSF andcomplementary products reduce complexity and management costs; facilitate better collaboration,
tighter integration and alignment of scheduling and resource management tasks with businessobjectives and priorities. LSF is architected to optimally place workloads not only based on the
capability of a cluster machine to run a workload, but based on a determination of what host is bestable to run the workload while ensuring broader business policies and requirements are met.
IBM Platform LSF lowers operating costs by smartly matching the limited supply of sharedresources with application demands and business priorities through features such as guaranteed
resources, live re-configuration, fair-share and pre-emptive scheduling enhancements, better performance and scalability. IBM continues to enhance the capabilities of LSF and LSF-add oncomponents. Clients can expect IBM to deliver capabilities to deploy new LSF add on componentson demand to keep up with ever changing requirements of the technical computing marketplace.
4 Trends from the trenches: Bio IT World 2012 http://www.slideshare.net/chrisdag/2012-trends-from-the-trenches
Technical
computing organizations
need
flexibility,
scalability,
and agility at
lowers costs
and risks
LSF can be
packaged with
engineering,
integration
and processes
so that
technical
computing
organizations
can become
more
productive and
focus on their
core business,
engineering or
scientific tasks
LSF lowers IT
costs smartly
by matching
the limited
supply of
shared
resources with
application
demands and
businesspriorities
LSF is tuned to
technical
computing now
and in the
future
7/29/2019 How the IBM Platform LSF Architecture Accelerates Technical Computing
http://slidepdf.com/reader/full/how-the-ibm-platform-lsf-architecture-accelerates-technical-computing 10/11
10
Appendix: What’s new in LSF Version 9.1?
IBM Platform LSF virtualizes heterogeneous IT infrastructure and offers customers complete freedom of
choice. Through fully integrated and certified applications, custom application integration, support for a wide
variety of operating systems, it ensures that current investments are preserved while providing the strategic
benefit of freedom of choice to run the best platform for the best job. The current LSF (Version 9.1) release
delivers improvements in performance and scalability over prior versions while introducing several additional
new features that simplify administration and boost productivity of cluster users.
New Featuresin LSF
Functional details Business Benefits Performanceand scalability
Improved Query Response ~10ms, decrease in Scheduling cycle, memoryoptimization, decrease start/restart time, parallel start-up/restart.
LSF has been extended to support an unparalleled scale of up to 160,000 cores
and two million queued jobs for very high throughput EDA workloads. On very large clusters with large numbers of user groups employing fair-share
scheduling, the memory footprint of the master batch scheduler in LSF has
been reduced by approximately 70% and scheduler cycle time has beenreduced by 25%.
Faster job turnaround times. Fastertime to results.
For a very large fair-share tree (e.g.
4K user group, 500 users with -g), jobelection performance has beenimproved 10x.
Better usability& manageability
Clearer reporting of resource usage and pending reasons. Better alternative job resource options for timely job execution
Enhanced process tracking: LSF 9.1 leverages kernel cgroup functionality to
replace/improve existing functionality for Process Tracking and Topology
CPU/memory enforcement. Fast detection of hung hosts/jobs, directory management. New multi-threaded
communication mechanism allows faster detection of unavailable hosts.
Speeds up troubleshooting, fasterdetection of failed or hung jobs, self-
tuning, and better admin productivity. Protection against user initiated
actions that can result in denial of service.
Timely job turnaround with alternateresources
Schedulingenhancements
LSF 8 provided guaranteed resource scheduling feature for groups of jobsbased on slots (cores), LSF 9.1 extended this feature for more complex
resource guarantees to support multi-dimensional packages. A package is acombination of slots and memory. This enables SLA scheduling to considermemory in addition to cores.
Besides numerous multi-cluster scheduling enhancements such as enhancedinteroperability across clusters and exchange of all load information between allclusters, it also provides CPU and memory affinity
LSF 9.1 also provides alternative or time based resource requirements to betteralign with business priorities with a much finer -grained control.
LSF scheduling enhancements makethe cluster more stable and reliable
Better job control and more accuratelight weight CPU and memoryaccounting even for run away and
short job processes. Fine-grained tuning and customization
of infrastructure sharing policies
ensure flexibility and agility inresource sharing that match closelywith evolving business requirements.
IBM Platform - AdvancedEdition Architecture
The new LSF - Advanced Edition architecture separates user interaction from
scheduling, and divides the compute resource into a number of executionclusters, while presenting it to the users as a single cluster.
This new architecture delivers the
expected increase in performance withthe increase in capacity resulting inconsistent user experience with scale.
LSF Add-on modules have also been enhanced in the latest version 9.1. The LSF License Scheduler handles parallel jobs where each rank checks out a license directly, more efficiently and does notneed a restart for making configuration changes. There are enhanced filtering and drill downcapabilities in IBM Platform RTM along with support for IBM General Parallel File System (GPFS)
monitoring. LSF Process Manager now supports non-LSF batch systems and the IBM PlatformSymphony product. IBM Platform Application Center has improved the interface with IBM PlatformAnalytics and the latter now supports Tableau (v8) and Vertica (5.1) and latest BI reportingcapabilities.
IBM Platform LSF V9.1 delivers significantly enhanced performance, scalability, manageability andusability as well as new scheduling capabilities. The new Platform LSF – Advanced Edition
provides greater than three times more scalability than prior versions of LSF, enabling clients to
consolidate their compute resources to achieve maximum flexibility and utilization.
For clients looking to improve service levels and utilization with a dynamic, shared HPC cloudenvironment, IBM Platform Dynamic Cluster V9.1 is now available as an add-on to IBM Platform
LSF. Platform Dynamic Cluster turns static Platform LSF clusters into dynamic, shared cloudinfrastructure. By automatically changing the composition of clusters to meet ever-changingworkload demands, service levels are improved and organizations can do more work with lessinfrastructure. With smart policies and numerous features such as live job migration and checkpoint-restart, Platform Dynamic Cluster enables clients to realize improved utilization, better reliability,
and increased productivity, while reducing administrator workload.
7/29/2019 How the IBM Platform LSF Architecture Accelerates Technical Computing
http://slidepdf.com/reader/full/how-the-ibm-platform-lsf-architecture-accelerates-technical-computing 11/11
11
The new IBM Platform Session Scheduler V9.1 is designed to work with Platform LSF to providehigh throughput, low-latency scheduling for a wide-range of workloads. It is particularly well suited
to environments that run high-volumes of short duration jobs, and where users require faster and more predictable job turnaround times. Unlike traditional batch schedulers that make resource allocation
decisions for every job submission, Platform Session Scheduler enables users to specify resourceallocation decisions only once for multiple jobs in a user session, providing users with their ownvirtual private cluster. With this more efficient scheduling model, users benefit from higher job
throughput and faster response times while cluster administrators realize an overall improvement incluster utilization.
To learn more about current IBM Platform LSF product features, visit:
http://www-03.ibm.com/systems/technicalcomputing/platformcomputing/products/lsf/index.html
Copyright ® 2012. Cabot Partners Group. Inc. All rights reserved. Other companies’ product names, trademarks, or service marks are used herein for identification only andbelong to their respective owner. All images and supporting data were obtained from IBM or from public sources. The information and product recommendations made by the Cabot Partners Group are based upon public information and sources and may also include personal opinions both of the Cabot Partners Group and others, all of which we believe to be accurate and reliable.However, as market conditions change and not within our control, the information and recommendations are made without warranty of any kind. The Cabot Partners Group, Inc. assumesno responsibility or liability for any damages whatsoever (including incidental, consequential or otherwise), caused by your use of, or reliance upon, t he information and
recommendations presented herein, nor for any inadvertent errors which may appear in this document. This document was developed with IBM funding. Although the document mayutilize publicly available material from various vendors, including IBM, it does not necessarily reflect the positions of such vendors on the issues addressed in this document.