Using XDMoD to Facilitate XSEDE Operations, Planning and Analysis
description
Transcript of Using XDMoD to Facilitate XSEDE Operations, Planning and Analysis
Using XDMoD to Facilitate XSEDEOperations, Planning and Analysis
Tom Furlani, PhDDirector - Center for Computational Research
University at Buffalo, SUNYXSEDE13 JULY 22 – 25, 2013
Thomas R. Furlani1, Barry I. Schneider2, Matthew D. Jones1, John Towns3, David L. Hart4, Steven M. Gallo1, Robert L. DeLeon1, Charng-Da Lu1, Amin Ghadersohi1, Ryan J. Gentner1,
Abani K. Patra5, Gregor von Laszewski6, Fugang Wang6, Jeffrey T. Palmer1, Nikolay Simakov1
1Center for Computational Research, University at Buffalo, SUNY, 2 CISE - Advanaced Computing Infrastructure, National Science Foundation, 3NCSA - University of Illinois,
4National Center for Atmospheric Research, 5Mech. & Aerospace. Eng. Dept. University at Buffalo, SUNY, 6Pervasive Technology Institute - University of Indiana
T E C H N O L O G Y A U D I T S E R V I C E
Outline• Overview of Technology Audit Service (XDMoD)• XDMoD Case Studies
– Data Driven CI Planning for XSEDE– System Operation and Maintenance– Interpreting XDMoD Data
• Future XDMoD Functionality– SUPReMM (Lightning Talk – Wed, 3PM, Marina Ballroom F&G)– PEAK (NICS) (Optimizing Utilization Across XSEDE – Thurs, 8:30AM, Marina Ballroom G)
– Scientific Impact and Open Source Version (XDMoD TAS BOF – Wed, 6PM, Palomar)
T E C H N O L O G Y A U D I T S E R V I C E
CoAuthors• Barry I. Schneider (NSF)• Matthew D. Jones (UB)• John Towns (NCSA)• David L. Hart (NCAR)• Steven M. Gallo (UB)• Robert L. DeLeon (UB) • Charng-Da Lu• Amin Ghadersohi (UB)• Ryan J. Gentner (UB)• Abani K. Patra (UB)• Gregor von Laszewski (Indiana)
• Fugang Wang (Indiana)• Jeffrey T. Palmer (UB)• Nikolay Simakov (UB)
T E C H N O L O G Y A U D I T S E R V I C E
Motivation• Measuring utilization of CI provides an understanding of how
resource is being utilized• HPC systems are a complex combination of software, processors,
memory, networks, and storage systems - difficult to know if optimal performance is being realized, or even if all subcomponents are functioning properly
0 200 400 600 800 10000
5,000,000
10,000,000
15,000,000
20,000,000
25,000,000
30,000,000
35,000,000
40,000,000Log Size As Of 9/12/2011
Node Number
Lo
g S
ize
(Byt
es)
job scheduler error node #126
loose cable node #348
Example: Log File Analysis Discovers Two Malfunctioning Nodes
T E C H N O L O G Y A U D I T S E R V I C E
XSEDE Technology Audit Service (TAS)• Provide Auditing and Quality of Service (QoS) Metrics• Primary components to TAS
– XDMoD: XSEDE Metrics on Demand Portal• Analytics Framework for XSEDE• Display results of all metrics (utilization, wait time, etc )• Easy to use
– Application Kernel Framework• Measure performance of XSEDE infrastructure• Diagnostic set of tools – early identification of system problems
• Broader Impact– Open source framework for academic HPC centers
• Organizations– Buffalo, Indiana (Laszewski), Michigan (Finholt), UT-NICS (You)
T E C H N O L O G Y A U D I T S E R V I C E
XDMoD Data Sources
T E C H N O L O G Y A U D I T S E R V I C E
XDMoD: XD Metrics on Demand Portal• Display metrics, Role Based, Custom Report Builder
T E C H N O L O G Y A U D I T S E R V I C E
XDMoD Case Studies
• Data Driven CI Planning for XSEDE• System Operation and Maintenance• Interpreting XDMoD Data
T E C H N O L O G Y A U D I T S E R V I C E
Data Driven CI Planning for XSEDE
• Largest, average and total SU allocations on XSEDE over time. Average and largest allocations have increased by more than a factor of 10 over the time period
9
T E C H N O L O G Y A U D I T S E R V I C E
Data Driven CI Planning for XSEDE• Total service unit usage by parent science- Molecular Bioscience usage has
grown over time – now rivals that of Physics
10
T E C H N O L O G Y A U D I T S E R V I C E
Data Driven CI Planning for XSEDE• However average core count varies widely over parent science – molecular
bioscience jobs tend to use a relatively small number of processors
11
T E C H N O L O G Y A U D I T S E R V I C E
CI System Operation and Maintenance • Application kernels help detect user environment anomaly at CCR• Example: Performance variation of NWChem due to bug in commercial parallel
file system that was subsequently fixed by vendor
T E C H N O L O G Y A U D I T S E R V I C E
CI System Operation and Maintenance • Sudden decrease in file system performance on TACC Lonestar4 as measured by 3
different application kernels (IOR, MPI-Tile-IO, and IMB)
T E C H N O L O G Y A U D I T S E R V I C E
CI System Operation and Maintenance • Application kernel control process to automatically detect underperforming
application kernels (poor performance). Red zone indicates an application kernel that is underperforming
T E C H N O L O G Y A U D I T S E R V I C E
Interpreting XDMoD Data• Like any analysis system, care must be exercised in interpretation of data
from XDMoD• Ex. Distribution of job sizes for all parent science Physics jobs in XSEDE
resources for the period 2008-2012
T E C H N O L O G Y A U D I T S E R V I C E
Interpreting XDMoD Data• Mean core count for Physics jobs in XSEDE resources for the period 2008-
2012, including (blue line) and excluding (red line) serial runs
High Throughput Jobs Start at Purdue
Number of Serial Physics Jobs by Resource
T E C H N O L O G Y A U D I T S E R V I C E
Future XDMoD Functionality: SUPReMM• SUPReMM (Lightning Talk – Wed, 3PM)
– Collaboration with TACC and U Texas at Austin– Comprehensive job level resource use measurement for large clusters – Will supply XDMoD with some missing job usage data – application run, memory,
local I/O, network, file-system, and CPU usage– Sample application report for Lonestar4
T E C H N O L O G Y A U D I T S E R V I C E
Future XDMoD Functionality: PEAK• NICS – PEAK (Thursday, 8:30AM)
– Optimizing Utilization Across XSEDE (Dr. Haihang You)– Performance Environment Autoconfiguration FrameworK– UT-NICS project to automatically tune key libraries and application kernels– Ex. Performance of Amber on Kraken – Amber built with PGI much faster
T E C H N O L O G Y A U D I T S E R V I C E
Future XDMoD FunctionalityOpen Source XDMoD & Scientific Impact
• Open Source Version: (XDMoD BOF - Wed, 6PM)– XDMoD functionality for non-XSEDE HPC centers– Installation by system administrators
• Programming not required• Guided textual installation process• Installation support provided by TAS Team
– Pre-existing central database not required• Aggregate data from available sources• Resource manager log files or existing database
– Currently recruiting for beta-testing program
• Scientific Impact– Preliminary XSEDE-based H-Index
T E C H N O L O G Y A U D I T S E R V I C E
Acknowledgement
• This work was sponsored by NSF under grant number OCI 1025159 for the development of Technology Audit Service for XSEDE.
• Contact Info– [email protected]– XDMoD https://xdmod.ccr.buffalo.edu/– [email protected]