Providing Scientific Software as a Service in Consideration of Service Level Agreements

Providing Scientific Software as a Servicein Consideration of Service Level Agreements

Oliver Niehörster(1), André Brinkmann(1), Georg Birkenheuer(1), Sonja Herres-Pawlis(2), Julia Niehörster(3), Jens Krüger(2), Brigitta Elsässer(2), Lars Packschies(4)

(1) Paderborn Center for Parallel Computing, Universität Paderborn, Germany

(2) Department Chemie, Universität Paderborn, Germany(3) Department Agricultural Sciences, Universität Hohenheim, Germany

(4) Universität zu Köln, Germany

19.04.23 1

Scientific SaaS

• Provider Advantages– Horizontal & vertical scaling of virtual machines – Better resource utilization– Snapshot & live migration– High availability, possible error recovery

19.04.23 2

CloudCloud

Job &

SLA

Job &

SLA

Service Stack

• Today: Concentration on business applications– Online word processors, content management, customer

relationship management, human resource management • Our contribution: Support for scientific applications

– Gaussian, Gromacs, MoE, NWChem, …

19.04.23 3

Agenda

19.04.23 4

MotivationArchitecture

Scientific SaaS Survey

Results

Architecture Stack

19.04.23 5

Optimal provisioning to fulfill SLAsFinding performance functions

and dynamic job scheduling

Scientific SaaS

Eucalyptus implements the EC2 APIConsideration of local resource management systems

Analysis of different mapping policies (Greedy, RR, Green-IT,...)

Cloud-/IaaS

Virtualization API to handle different hypervisorsExtensions of libvirt to handle Hyper-V and ESX

Virtualization

High-performance interconnects (like Infiniband)Analysis of VMM-bypass, MR-IOV, SR-IOV

HPC Hardware

Is this a

pproach applicable to scientific

applications?

How do scientific applications behave?

Is this a

pproach applicable to scientific

applications?

How do scientific applications behave?

Aim of the survey

• Can classical SLA used to provide information about scientific applications?– Can we estimate temporal behaviour of the

applications?– Can we support disaster recovery?

• Can we use virtualisation capabilities to increase scheduling efficiency?– Resizing of virtual machines? – Change number of nodes?

19.04.23 6

Questionnaire

19.04.23 7

• Durable service: sustainable service runs 24/7 • Batch job: one-time computations (always halts)• Experiment: unknown termination behavior

Activity categories

• Live progress status• Benchmark• Estimation function• Unknown

Method to measure performance

Result

19.04.23 8

Kind/Method Live progress status

Benchmark Estimation Function

Unknown

Durable Service

Batch Job Plabsoft, R, Gromacs,NWChem

ScientificGromacs

Gaussian

Experiment Gromacs Gromacs Gaussian SAS, ASRemlScientificScientific

BusinessBusiness

• How does application scale (max. CPUs, etc)?– Parallelism is user decision - up to 1024 CPUs

• How does parallelism behave over time?– no change in parallelism

• Is it possible to add nodes on line?– No!

• Is the resource demand time dependent?– Application dependent: constant or increasing over

time• High internode or I/O communication?– All computing intensive, most I/O intensive

Application Parallelism

19.04.23 9

Application Input

• Are interactive user inputs possible during calculation?– No

• Is all input available initial or are workflow dependencies? – Initial available

• How much Data uses the application?– Several MB to several GB

19.04.23 10

Application Progress Indication

• Does the application provide progress information?– Progress often not available – If available• Gromacs reliable after 1000 iterations• R -> not linear• PlabSoft, SAS, ASRemL -> unreliable

• Guess end time from history

19.04.23 11

Checkpoint and Restart

• Is application specific checkpointing possible?– Unknown for some applications• PlabSoft, SAS, ASRemL

– Manually started • Gromacs

– Automatically started• R, Gaussian, NWChem

• Can a checkpoint continue computation with more nodes?– If available then possible

19.04.23 12

Summary & Conclusion

• Summary– Scientific applications are batch jobs– Halting problem: Determination of application finish

not possible in every case– No online progress indication– Virtualisation

• Online horizontal scaling virtual machines not supported • Vertical scaling helps

• Conclusion– Providing scientific SaaS is challenging– Limited support for SLAs

19.04.23 13

Georg [email protected]

19.04.23 14