Informationsaustausch für Nutzer des Aachener HPC Clusters



Similar documents
HPC-Nutzer Informationsaustausch. The Workload Management System LSF

RWTH GPU Cluster. Sandra Wienke November Rechen- und Kommunikationszentrum (RZ) Fotos: Christian Iwainsky

The RWTH Compute Cluster Environment

Using the Windows Cluster

How To Build A Supermicro Computer With A 32 Core Power Core (Powerpc) And A 32-Core (Powerpc) (Powerpowerpter) (I386) (Amd) (Microcore) (Supermicro) (

Introduction to Running Computations on the High Performance Clusters at the Center for Computational Research

Parallel Processing using the LOTUS cluster

Virtualization of a Cluster Batch System

Work Environment. David Tur HPC Expert. HPC Users Training September, 18th 2015

An HPC Application Deployment Model on Azure Cloud for SMEs

Debugging with TotalView

Cloud Computing through Virtualization and HPC technologies

Visualization Cluster Getting Started

OpenMP Programming on ScaleMP

SLURM Workload Manager

Eine CAE Infrastruktur für LS-DYNA. unter Verwendung von. Microsoft Windows HPC Server 2008

Matlab on a Supercomputer

UMass High Performance Computing Center

Streamline Computing Linux Cluster User Training. ( Nottingham University)

Chapter 2: Getting Started

HPC at IU Overview. Abhinav Thota Research Technologies Indiana University

PEPPERDATA IN MULTI-TENANT ENVIRONMENTS

ABAQUS High Performance Computing Environment at Nokia

PARALLELS SERVER BARE METAL 5.0 README

A High Performance Computing Scheduling and Resource Management Primer

Java Bit Torrent Client

Agenda. Using HPC Wales 2

Windows HPC 2008 Cluster Launch

White Paper Perceived Performance Tuning a system for what really matters

High Performance Computing in Aachen

Recent Advances in HPC for Structural Mechanics Simulations

Using the Yale HPC Clusters

The Asterope compute cluster

An Introduction to High Performance Computing in the Department

- An Essential Building Block for Stable and Reliable Compute Clusters

High Performance Computing within the AHRP

Integration of Virtualized Workernodes in Batch Queueing Systems The ViBatch Concept

LS-DYNA Best-Practices: Networking, MPI and Parallel File System Effect on LS-DYNA Performance

Energy-aware Memory Management through Database Buffer Control

Parallel Computing with MATLAB

High Performance Computing in CST STUDIO SUITE

Tableau Server 7.0 scalability

Lecture 2 Parallel Programming Platforms

Parallels Plesk Panel

Achieving Nanosecond Latency Between Applications with IPC Shared Memory Messaging

Using WestGrid. Patrick Mann, Manager, Technical Operations Jan.15, 2014

PES. Batch virtualization and Cloud computing. Part 1: Batch virtualization. Batch virtualization and Cloud computing

Getting Started with HPC

The PHI solution. Fujitsu Industry Ready Intel XEON-PHI based solution. SC Denver

System Requirements Table of contents

IMPLEMENTING GREEN IT

MFCF Grad Session 2015

Improved metrics collection and correlation for the CERN cloud storage test framework

Adaptive Resource Optimizer For Optimal High Performance Compute Resource Utilization

ECDF Infrastructure Refresh - Requirements Consultation Document

OIS. Update on Windows 7 at CERN & Remote Desktop Gateway. Operating Systems & Information Services CERN IT-OIS

Hodor and Bran - Job Scheduling and PBS Scripts

Mark Bennett. Search and the Virtual Machine

Monitoring Databases on VMware

Signiant Agent installation

Moab and TORQUE Highlights CUG 2015

High Productivity Computing With Windows

Parallel Programming Survey

Grid Engine Basics. Table of Contents. Grid Engine Basics Version 1. (Formerly: Sun Grid Engine)

Running COMSOL in parallel

Microsoft HPC. V 1.0 José M. Cámara (checam@ubu.es)

Introduction 1 Performance on Hosted Server 1. Benchmarks 2. System Requirements 7 Load Balancing 7

Using esxtop to Troubleshoot Performance Problems

Parallel Computing using MATLAB Distributed Compute Server ZORRO HPC

OLCF Best Practices. Bill Renaud OLCF User Assistance Group

Tableau Server Scalability Explained

Large system usage HOW TO. George Magklaras PhD Biotek/NCMM IT USIT Research Computing Services

SLURM: Resource Management and Job Scheduling Software. Advanced Computing Center for Research and Education

Quick Start Guide for Parallels Virtuozzo

Scyld Cloud Manager User Guide

CHAPTER FIVE RESULT ANALYSIS

Using NeSI HPC Resources. NeSI Computational Science Team

SOSFTP Managed File Transfer

MATLAB Distributed Computing Server System Administrator's Guide

SYSTEM SETUP FOR SPE PLATFORMS

HPC Cloud. Focus on your research. Floris Sluiter Project leader SARA

An introduction to compute resources in Biostatistics. Chris Scheller

IBM LoadLeveler for Linux delivers job scheduling for IBM pseries and IBM xseries platforms running Linux

WA1791 Designing and Developing Secure Web Services. Classroom Setup Guide. Web Age Solutions Inc. Web Age Solutions Inc. 1

Capacity Planning Process Estimating the load Initial configuration

Windows Server 2008 R2 Essentials

Optimizing Shared Resource Contention in HPC Clusters

Introduction to Supercomputing with Janus

ZEN LOAD BALANCER EE v3.02 DATASHEET The Load Balancing made easy

The CNMS Computer Cluster

Guideline for stresstest Page 1 of 6. Stress test

Best Practices for Monitoring Databases on VMware. Dean Richards Senior DBA, Confio Software

Efficient Load Balancing using VM Migration by QEMU-KVM

Introduction to Linux and Cluster Basics for the CCR General Computing Cluster

IN STA LLIN G A VA LA N C HE REMOTE C O N TROL 4. 1

High Performance Computing in Aachen

Computing in High- Energy-Physics: How Virtualization meets the Grid

PARALLELS SERVER 4 BARE METAL README

Using SAS in a Distributed Computing Environment Chris Smith, Platform Computing, San Jose, CA

Assignment # 1 (Cloud Computing Security)

Transcription:

Informationsaustausch für Nutzer des Aachener HPC Clusters Paul Kapinos, Marcus Wagner - 21.05.2015

Informationsaustausch für Nutzer des Aachener HPC Clusters Agenda (The RWTH Compute cluster) Project-based management of the cluster resources Interactive usage Using the batch system Integratives Hosting Discussion 3 von 5

The RWTH Compute cluster No. 32 in TOP500 (June 2011), no. 4 in Germany No. 272 in TOP500 (Nov 2014) MPI Partition 1358 Westmere EP nodes 2x Xeon X5675 (6-core CPU) @ 3.06 GHz => 16k cores 24 or 96 GB RAM (4:1) SMP Partition 88 Nehalem EX nodes 16x Xeon X7550 (8-core CPU) @ 2.00 GHz => 11k cores 256 2048 GB RAM Interactive front ends and back ends ( 1% of cluster) 300 cores, max. 256 GB RAM, 100++ users p.n. (GPU cluster, MIC cluster, old hardware, IH systems)

Why? Because we must: requirements of the Science Council (Wissenschaftsrat) to request funding for future computer systems:. a scientific process for the allocation of the expensive compute resources has to be established which will guarantee fair handling of all users.

Why? (ii) fair distribution of resources main goals: collation between used resources and scientific value defined (short...) job staring times defined, predictable throughput for researchers effective and resource saving usage pattern would you drive fuel-saving, if you do not have to pay for the fuel? if you indeed have no clue how many fuel did you burned? and, last but not least, to keep some buddies within bounds

The status Implemented with projects and queues in LSF batch system JARA-HPC partition (30% ): since 2012 general introduction: Q3/2014 up n runnin now Use a project: add a line to your batch file #BSUB -P abcd4321 Check your quota: $ r_batch_usage

How? (II) Free quota scientific employees: 2000 core-h per month (about a week of 12-core node p.m.) students: 500 core-h per month Need more? file a project! RWTH Small (S): up to 10,000 core-h p.m. (0.01 mio core-h p.m.) technical review only RWTH Standard (M): up to 0.05 mio core-h p.m. a project description is required (=> internal scientific review) JARA-HPC/RWTH Big (XL/L): up to 2.5 mio core-h p.m. submission twice a year following the JARA-HPC procedures a detailed project description is required (=> external scientific reviews) student need more for lecture / course / thesis? RWTH lecture, RWTH thesis (XS) (up to 20,000 core-h p.m. (*) ) technical review only

How-to file an Application for computing time go to https://doc.itc.rwth-aachen.de/display/cc/projektbewirtschaftung decide for what type of project you should apply determine your needs; don t be shy! don t try to be too exact It s better to ask for 30% too much than 1% too less it s easier to ask for round sum (compute time, duration..) both for you and us think about special requirements: overlong compute time? (more that 120h not possible) disk storage? one huge project, or maybe multiple subprojects?.

How-to file an Application for computing time (II) go to https://doc.itc.rwth-aachen.de/display/cc/projektbewirtschaftung fill in the right form use Acrobat X to edit the PDF file we need the data to be extracted electronically do not use meaningless values like normal, much for e.g. memory consumption do not cut a corner we do not know who Mr See Above is! send the electronically-readable PDF file to servicedesk@itc.rwth-aachen.de do not send us screen shots, JPG, PNG, DOCX, TXT files do not send us signed+scanned PDFs via email print the same file, sign it, and fax or mail it to us do not send send us signed+scanned PDFs via email At the end we need the same document in two versions: signed+legal (thus fax or mail), and electronically readable.

How-to file an Application for computing time (III) go to https://doc.itc.rwth-aachen.de/display/cc/projektbewirtschaftung filing an application for RWTH Standard (M) project? A project description is required (for internal scientific reviews) Bring up if your project is a follow-up project, is funded by some organisation, filing an application for JARA-HPC/RWTH-Big (XL/L) project? Submission twice a year following the JARA-HPC procedures http://www.jara.org/index.php?id=565&s=1&l=2

How-to file an Application for computing time (IV) go to https://doc.itc.rwth-aachen.de/display/cc/projektbewirtschaftung Application form filled, emailed, printed out, signed, faxed/mailed? then wait In typically a week: a message that both versions of application form has been arrived. Some days later: 1) a message that the project is ready-to-use, (for small projects), or 2) a message that the project has been introduced with a test quota of 0.01 Mio corehours per month, and the scientific review process started (for larger projects) For (2), some weeks (or either months ) later: a message that the project is approved and full remaining quota is granted (often the runtime of project is adjusted, too, according to the delay) Yes we know: this process is a really tedious and lengthy one Working on improving it. But at least the scientific review will stay delaying factor.

What happen if over quota? running jobs continue to the end (and still consume core-h!) new-submitted and pending jobs moved to low-priority queue they still can start! but if and only if there are free resources not used by normal-priority jobs if started from low queue, still consume core-h quota may go well in the red! today, no hard limit in low-priority queue this will be subject of change in future, very likely at 1 st of any month, you get next month s quota added. if you are in the black with your quota then, new jobs will be submitted and pending jobs will be moved to normal-priority queue technically, no difference from which queue job is started only start time differs!

How is the quota computed? main goal is to motivate the users for continuous using of resources but still allow some peaks Three-month sliding window up to 300% of month quota available in a month unused quota from previous month is transferred to the current month but not further The quota for the previous, the current, and the next month are added up The consumed core-h for the previous and for the current month are added up The difference between both values is the amount of core-h available in the current month Huh? https://doc.itc.rwth-aachen.de/display/cc/resource+contingents

Check your quota now! Check your quota: $ r_batch_usage h (manual of r_batch_usage) $ r_batch_usage (overwiew; big terminal window advisable) $ r_batch_usage -q User: pk224850 Status of user: RWTH-Mitarbeiter Quota monthly (core-h): 2000 Remaining core-h of prev. month: -4000 Consumed core-h act. month: 3938 Consumable core-h (%): -101 Consumable core-h: 0

Check your project quota now! Usage of batch system with projects: $ r_batch_usage h (manual of r_batch_usage) $ r_batch_usage p <projectname> (overwiew) $ r_batch_usage -p jara0001 -q Group: jara0001 Start of Accounting Period: 01.05.2015 End of Accounting Period: 01.05.2016 State of project: active Quota monthly (core-h): 100000 Remaining core-h of prev. month: 0 Consumed core-h act. month: 30331 Consumable core-h (%): 70 Consumable core-h: 169669

Informationsaustausch für Nutzer des Aachener HPC Clusters Agenda (The RWTH Compute cluster) Project-based management of the cluster resources Interactive usage Using the batch system Integratives Hosting Discussion 3 von 5

Interactive usage Batch system: MPI Partition 1358 Westmere EP nodes => 16k cores SMP Partition 88 Nehalem EX nodes=> 11k cores Interactive front ends and back ends ( 1% of cluster) 300 cores, max. 256 GB RAM, 100++ users p.n. challenged resource! an issue with an interactive node directly affect 100s of users we want to enable as much of test options as possible notorious trade-off between feature and stability

Interactive usage Go to: https://doc.itc.rwth-aachen.de/display/cc/interactive+usage Interactive front ends cluster.rz.rwth-aachen.de, cluster-linux.rz.rwth-aachen.de (main front ends) cluster-copy.rz.rwth-aachen.de, cluster-copy2.rz.rwth-aachen.de (for file transfer only) cluster-x.rz.rwth-aachen.de, cluster-x2.rz.rwth-aachen.de (GUI / remote desktop). and others. supported protocols: SSH (with X11 forwarding), SCP, remote desktop (FastX/XWin32) $ ssh -X -l ab123456 cluster.rz.rwth-aachen.de accessible from RWTH network only (VPN helps!) Interactive back ends used to off-load MPI processes started on front ends off-loading managed by Interactive MPIEXEC wrapper hardware subject of change currently: 8x 12-core Westmere with 96 GB RAM 3 von 5

Interactive usage Go to: https://doc.itc.rwth-aachen.de/display/cc/interactive+usage Interactive front ends are frequented by 100s of users! any issue directly interrupt work of these users! Purposes: data transfer, job submission, application porting, testing, tuning, debugging NOT FOR PRODUCTIVE RUNS USE BATCH SYSTEM Rule of thumb: not more than 20 minutes of CPU time that does not mean I can start 80x of 19.5-minute-runs one after another! Really need compute power and interactive session? Batch jobs with GUI: https://doc.itc.rwth-aachen.de/display/cc/submitting+a+job+with+gui In terms of advanced testing, we set flexible quotas using cgroup system CPU: processes of a user are configured to get the same amount of CPU cycles as all processes of other user Memory: real memory is limited to a part of available RAM - this prevent the situation when one user consume all RAM and crash the whole node - use memquota command to find out current situation 3 von 5

Interactive usage Go to: https://doc.itc.rwth-aachen.de/display/cc/interactive+usage Interactive front ends are frequented by 100s of users! any issue directly interrupt work of these users! Main Performance Issue on front ends: DATA TRANSFER Use dedicated front ends for any data transfer, TAR, ZIP, cluster-copy.rz.rwth-aachen.de cluster-copy2.rz.rwth-aachen.de 3 von 5

Interactive usage Go to: https://doc.itc.rwth-aachen.de/display/cc/interactive+usage Interactive back ends unprovided for log in hardware subject of change (currently: 8x 12-core Westmere with 96 GB RAM) used to off-load MPI processes started on front ends (reduce load!) off-loading managed by Interactive MPIEXEC wrapper example: $MPIEXEC np 2 hostname processes started on less-loaded nodes, but with massive overloading allowed further, you re not alone on these systems load of 100+ is not unusual any productive runs and time measurements absurd to the highest degree the only sense of (overloading) test runs: test of will my binary start with XYZ ranks? - if yes, Ctrl-C and proceed to the Batch System - if not, you ve got the reply immediately (instead of waiting a day for the batch job) NOT FOR PRODUCTIVE RUNS USE BATCH SYSTEM https://doc.itc.rwth-aachen.de/display/cc/testing+of+mpi+jobs 3 von 5

Interactive usage 2014-11-14: Changed Terms of Use passing along your HPC account to third parties is explicitly forbidden secondary logins will be gradually deactivated during the next months secondary accounts (after 05/2014) already configured without login permissions the only use of secondary accounts now: data sharing Jobs and processes (in Batch, on interactive front ends and back ends) which disturb other jobs/processes, may be killed without further notice. If your job has been killed, you probably do some bad thing read the documentation! https://doc.itc.rwth-aachen.de/display/cc/2014/12/02/2014-11- 14%3A+Changed+Terms+of+Use

"Would You Like to Know More?" Links The Primer, It s worth reading: http://www.itc.rwth-aachen.de/hpc/primer slightly outdated now update planned The Dokuweb, it s most actual and worth reading, too: Linux Cluster: https://doc.itc.rwth-aachen.de/display/cc/home Parallel Programming: https://doc.itc.rwth-aachen.de/display/ccp/home Events: https://doc.itc.rwth-aachen.de/display/ve/home Today s themes in Dokuweb: Project-based management of the cluster resources https://doc.itc.rwth-aachen.de/display/cc/projektbewirtschaftung Manuals: https://doc.itc.rwth-aachen.de/display/cc/manuals Interactive usage: https://doc.itc.rwth-aachen.de/display/cc/interactive+usage Batch system: https://doc.itc.rwth-aachen.de/display/cc/using+the+batch+system