dynamic adaptivity in support of extreme scale pat teller, utep
DESCRIPTION
FastOS. Dynamic Adaptivity in Support of Extreme Scale Pat Teller, UTEP. Outline Collaborators Overview Progress Plans. Partners. University of Texas at El Paso Department of Computer Science Patricia J. Teller ([email protected]). University of Wisconsin — Madison - PowerPoint PPT PresentationTRANSCRIPT
9 June 2005 FastOS PI Meeting 1
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Dynamic Adaptivity
in Support of Extreme Scale
Pat Teller, UTEP
9 June 2005 FastOS PI Meeting 2
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Outline
• Collaborators
• Overview
• Progress
• Plans
9 June 2005 FastOS PI Meeting 3
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
University of Texas at El PasoDepartment of Computer SciencePatricia J. Teller ([email protected])
University of Wisconsin — MadisonComputer Sciences DepartmentBarton P. Miller ([email protected])
International Business Machines, Inc.Linux Technologies CenterBill Buros ([email protected])
U.S. Department of EnergyOffice of ScienceFred Johnson ([email protected])
Partners
Lawrence Berkeley National Laboratory Leonid Oliker ([email protected])new partner
9 June 2005 FastOS PI Meeting 4
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
• UTEP team (Pat Teller)– Rodrigo Romero, Ph.D. (Post-doc)– Seetharami Seelam, Ph.D. candidate in CS– Luis Ortiz, Ph.D. candidate in CS– Jayaraman Suresh, Master’s candidate in CS– Brenda Prieto, Master’s candidate in ECE– Nidia Pedregon, Undergraduate in CS– Alejandro Castaneda, Undergraduate in CS
• U. Wisconsin-Madison team (Bart Miller)– Michael Brim, Ph.D. candidate– Igor Grobman, Ph.D. candidate
Teams
9 June 2005 FastOS PI Meeting 5
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Outline
• Collaborators
• Overview– Goal– Challenges– Deliverables– Methodology
• Progress & Direction
9 June 2005 FastOS PI Meeting 6
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Goal:Goal: Enhanced PerformanceEnhanced Performance
Generalized CustomizedCustomized resource management
Fixed Dynamically AdaptableDynamically Adaptable OS/runtime services
9 June 2005 FastOS PI Meeting 7
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Determining
• WhatWhat to adapt
• WhenWhen to adapt
• HowHow to adapt
• HowHow to measure effects of adaptation
Challenges
9 June 2005 FastOS PI Meeting 8
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
1. Mechanisms to dynamically sense, analyze, and adjust
• performance metrics• fluctuating workload situations • overall system environment
conditions
Deliverables
9 June 2005 FastOS PI Meeting 9
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
2. Linux prototypes and experiments that demonstrate dynamic self-tuning / provisioning in HPC environments
Deliverables
9 June 2005 FastOS PI Meeting 10
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
3. Methodology for general-purpose OS adaptation
Deliverables
9 June 2005 FastOS PI Meeting 11
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
identify adaptationtargets
characterize workloadresource usage patterns
determine/redetermine feasible adaptation ranges
define/adapt metrics/heuristics to trigger adaptation
generate/adapt monitoring, triggering andadaptation code, and attach it to OS
potentially profitable adaptation targets
Methodology
KernInstmonitor application execution,
assessing performance (gain) and triggering adaptation as necessary
off line
off line/run time
9 June 2005 FastOS PI Meeting 12
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
InstrumentationTool
Client
KernInst APIKernInst Device
Linux Kernel
KernInst Daemon
IBM pSeries eServer 690
KernInst
dynamic monitoring, instrumentation, and adaptation of the kernel
9 June 2005 FastOS PI Meeting 13
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Outline
• Collaborators
• Overview
• Progress & Direction– Tools– Infrastructure– Collaboration– Research
9 June 2005 FastOS PI Meeting 14
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Tools Progress• KernInst –
– POWER4 port for Linux 2.4 and IA32 Linux 2.6• Modifications/Enhancements for DAiSES research
– POWER4 port for Linux 2.6 underway
• IOstat (coarse statistics) – POWER4 Linux 2.6
• Investigating complementary use of oprofile and kprobes – POWER4 Linux 2.6
9 June 2005 FastOS PI Meeting 15
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
KernInst• Intel IA32 port for Linux 2.4 and 2.6 for Pentium 3 and Pentium 4 processors • IBM POWER4 port for Linux 2.4• Supports stand-alone kernels and kernels that run under the Hypervisor virtual machine
layer• Hypervisor layer not transparent and requires explicit support• Little public documentation on this layer available
9 June 2005 FastOS PI Meeting 16
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Outline
• Collaborators
• Overview
• Progress & Direction– Tools– Infrastructure– Collaboration– Research
9 June 2005 FastOS PI Meeting 17
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Infrastructure ProgressIBM SUR Grants, UTEP Star Award, UTEP PUF Funds
• Development of Experimental Platforms at UTEP– IBM eServer pSeries 690 (16 processors, 32GB, 2TB)
• Linux 2.4/2.6 partition for KernInst development• Linux 2.4 partition for DAiSES research• Linux 2.6 partition for DAiSES research• DS4300 RAID for DAiSES I/O-related research (1TB)
– Xeon workstations – Linux 2.4 and 2.6– IBM eServer p590 (24 processors, 64GB, 2TB)– IBM eServer p550 (4 processors, )
• Establishment of DAiSES Lab at UTEP
9 June 2005 FastOS PI Meeting 18
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Outline
• Collaborators
• Overview
• Progress & Direction– Tools– Infrastructure– Collaboration– Research
9 June 2005 FastOS PI Meeting 19
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Build/Strengthen Collaborations - 1• April 14-16, 2004 – Seetharami Seelam, UTEP, attended
Paradyn/Condor week• October 24, 2004 – Barney MacCabe, UNM, visited UTEP to meet
with the DAiSES team and give a talk re: his team’s FASTOS research
• November 10, 2004 – Pat Teller, UTEP, participated in the FASTOS Birds-of-a-Feather meeting at SC2004 – this was the first public presentation of the DAiSES project
• November 2004 – Rodrigo Romero, Seetharami Seelam, and Pat Teller, UTEP, and Michael Brim, Igor Grobman, and Bart Miller, UW-Madison, promoted the DAiSES project at two SC2004 research exhibits, one shared by UTEP, UNM, New Mexico State University, and New Mexico Institute of Technology, and another of UW-Madison
9 June 2005 FastOS PI Meeting 20
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Build/Strengthen Collaborations - 2• February 25, 2005 – Luis Ortiz, Rodrigo Romero, Seetharami
Seelam, and Pat Teller, UTEP, attended a half-day meeting at IBM-Austin with approximately nine members of the Linux Technologies team
• March 3, 2005 –Rodrigo Romero, Seetharami Seelam, and Pat Teller, UTEP, attended an all-day meeting at UNM with Barney Maccabe, Patrick Bridges, Kurt Ferreira, an Edgar Leon, UNM/CS, Orran Krieger, IBM, Ron Brightwell and Rolf Riesen, SNL, and Rod Oldehoeft, LANL/ACL
• March 20-24, 2005 – Igor Grobman, UW-Madison, visited UTEP and led a workshop re: the use of KernInst and Kperfmon to implement adaptations, in particular, in the process scheduler and I/O scheduler; Bill Buros, IBM-Austin, senior member of Linux Technologies team, attended for three days – resulted in modifications/enhancements to KernInst
9 June 2005 FastOS PI Meeting 21
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Build/Strengthen Collaborations - 3• May 4, 2005 – after IBM Petaflops Tools Strategy Workshop at IBM
TJ Watson Research Center, Bart Miller, UW-Madison, and Seetharami Seelam (awarded $500 to attend the workshop) and Pat Teller, UTEP, met with Evelyn Duesterwald and Robert Wisniewski re: possible collaborations
• Weekly telecons with IBM-Austin team• Telecons and Access Grid meetings with UW-Madison team• Shared Enotebook to be launched shortly• Shared data repository with search tool to be launched shortly
9 June 2005 FastOS PI Meeting 22
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Outline
• Collaborators
• Overview
• Progress & Direction– Tools– Infrastructure– Collaboration– Research
9 June 2005 FastOS PI Meeting 23
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Current Research Thrusts
• Dynamic Code Optimization• Low-hanging Fruit (i.e., opportunitistic
targets of adaptation)• Identification of Adaptation Targets via Self-
propelled Instrumentation• Other Directions
9 June 2005 FastOS PI Meeting 24
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Dynamic Code Optimization• Investigation of dynamic code optimization strategies
– [Tamches and Miller] used dynamic reorganization of basic block layout in parts of SPARC Solaris kernel to improve performance via I-cache miss reduction
• Discussion in research community asks if such optimizations are – workload dependent and need to be done dynamically
or – mostly independent of workload and can be done
statically• Goal: to provide conclusive evidence either way
9 June 2005 FastOS PI Meeting 25
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Low-hanging Fruit
I/O scheduling – extend work of IBM-Austin
process scheduling – extend published observations and address daemon controlvirtual memory management –
extend dissertation work page size – extend work of IBM TJ Watson Research Ctr.
I/O scheduler parameter selection via neural networks – extend work of IBM-Austin that uses genetic algorithms
9 June 2005 FastOS PI Meeting 26
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
I/O Scheduling (in progress)• WhatWhat to adapt: I/O scheduler
• Dynamic selection of “appropriate” I/O scheduler for observed “system state”
• WhenWhen to adapt• Change in “system state” (now identified via IOstat)• Below threshold related to number of queued I/O
requests• HowHow to adapt
• Linux 2.6 provides capability• I/O schedulers characterized w.r.t. “best” performance
for different “system states” [Pratt and Heger]• HowHow to measure effects of adaptation
• Execution time and throughput MB/s (for now)
9 June 2005 FastOS PI Meeting 27
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
I/O Scheduling By-products - 1
Enhancements to Linux I/O Scheduling,” to appearEnhancements to Linux I/O Scheduling,” to appearin in Proceedings of the Linux SymposiumProceedings of the Linux Symposium, Ottawa, , Ottawa, Canada, July 2005 (S. Seelam, R. Romero, P. Canada, July 2005 (S. Seelam, R. Romero, P. Teller, and W. Buros)Teller, and W. Buros)
• Reviews previous work of IBM-Austin characterizes Reviews previous work of IBM-Austin characterizes workloads best served by each of the four Linux 2.6 I/O workloads best served by each of the four Linux 2.6 I/O schedulers (can be selected at boottime or runtime)schedulers (can be selected at boottime or runtime)
• Presents cases where the Anticipatory Scheduler (AS) Presents cases where the Anticipatory Scheduler (AS) results in process starvation results in process starvation
9 June 2005 FastOS PI Meeting 28
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
I/O Scheduling By-products - 1 cont’d.
• Presents and demonstrates performance of UTEP CAS, Cooperative Anticipatory SchedulerCooperative Anticipatory Scheduler• extends anticipation to “cooperative” processes that extends anticipation to “cooperative” processes that
collectively issue synchronous requests to a close set collectively issue synchronous requests to a close set of disk blocksof disk blocks
• compares performance to current four schedulers compares performance to current four schedulers • shows order of magnitude performance improvement shows order of magnitude performance improvement
in cases where AS performs poorlyin cases where AS performs poorly
9 June 2005 FastOS PI Meeting 29
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
I/O Scheduling By-products - 2
In progress: demonstration of heuristically-guideddynamic selection of Linux 2.6 I/O schedulers (target: FAST)
• First step towards making I/O scheduling fully autonomic First step towards making I/O scheduling fully autonomic (Ph.D. dissertation topic: Seetharami Seelam)(Ph.D. dissertation topic: Seetharami Seelam)
• Selection based on observed system behavior, i.e., Selection based on observed system behavior, i.e., system (workload) I/O behavior, metric, in particular, I/O system (workload) I/O behavior, metric, in particular, I/O request sizerequest size
9 June 2005 FastOS PI Meeting 30
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
I/O Scheduling By-products - 2 cont’d.
• Using Using a prioria priori measurements of disk throughput under measurements of disk throughput under the various schedulers and request sizes to generate a the various schedulers and request sizes to generate a function that at runtime, given the current average function that at runtime, given the current average request size, returns the scheduler that gives the best request size, returns the scheduler that gives the best measured throughput for the specified diskmeasured throughput for the specified disk
• Identying adaptation interval, i.e., when it is not too Identying adaptation interval, i.e., when it is not too expensive to switch schedulers–based on number of expensive to switch schedulers–based on number of queued I/O requestsqueued I/O requests
• Future work: UW-Madison team will use KernInst to Future work: UW-Madison team will use KernInst to effect the adaptationeffect the adaptation
9 June 2005 FastOS PI Meeting 31
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
0
2
4
6
8
10
12
14
1 2 4 8 16 32 64 128 256 512 1024
Number of Requests
Ove
rhea
d (1
000
ms)
.
Deadline to AS
AS to Deadline
Time to Drain for Default nr_requests
Overhead of Draining I/O Queue
9 June 2005 FastOS PI Meeting 32
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Comparison of Different I/O Schedulers on RAID-0
0
9
18
27
36
45
1KB 4KB 16KB 32KB 65KB 256KB 1MB
Read Size
Ban
dw
idth
(M
B/s
)
. CFQ
AS
Deadline
NOOPPoints of interests
AS
Deadline
CFQ
Microbenchmark Synchronous Reads
9 June 2005 FastOS PI Meeting 33
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Comparison of Different I/O Schedulers on RAID-0
1
2
3
4
5
6
1KB 4KB 16KB 32KBRead Size
Ban
dw
idth
(MB
/s) .
CFQ
AS
Deadline
NOOP
30% difference between the best and the worst performing scheduler
Microbenchmark Synchronous Reads ZOOMED 1
9 June 2005 FastOS PI Meeting 34
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Comparison of Different I/O Schedulers on RAID-0
0
5
10
15
20
25
30
35
40
45
50
32KB 65KB 256KB 1MB
Read Size
Bandw
idth
(M
B/s
) .
CFQ
AS
Deadline
NOOP
>20% difference between the best and the
worst performing scheduler
Microbenchmark Synchronous Reads ZOOMED 2
9 June 2005 FastOS PI Meeting 35
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Comparison of Different I/O Schedulers on RAID-0
0
9
18
27
36
45
54
1KB 4KB 16KB 64KB 256KB 1MB
Read Size
Ban
dw
idth
(M
B/s
) .
CFQ
AS
Deadline
NOOP
MicrobenchmarkWrites
9 June 2005 FastOS PI Meeting 36
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Linux Compilation Disk Accesses
Histogram of Size of Linux Object Files
0
500
1000
1500
2000
2500
Size
Nu
mb
er o
f F
iles
.
Histogram of the Size of Linux Source files
0
1000
2000
3000
4000
5000
6000
0K 1K 2K 4K 8K 16K 32K 64K 128K 256K 512K 1M 2M
Size
Nu
mb
er o
f F
iles
.
Comparison of Different I/O Schedulers on RAID-0
0
9
18
27
36
45
54
1KB 4KB 16KB 65KB 256KB 1MB
Read Size
Ban
dw
idth
(M
B/s
)
.
CFQ
AS
Deadline
NOOP
Points of interests
Deadline
CFQ
Comparison of Different I/O Schedulers on RAID-0
0
9
18
27
36
45
1KB 4KB 16KB 32KB 65KB 256KB 1MB
Read Size
Ban
dw
idth
(M
B/s
)
. CFQ
AS
Deadline
NOOPPoints of interests
AS
Deadline
CFQ
9 June 2005 FastOS PI Meeting 37
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
• Dynamic Adaptation– Uses IOstat information, which is after the fact– Read/Write prediction uses two-bit saturating
counter– For reads: size 1K - 32K use Deadline
size > 32K use AS
– For writes: size 1K - 64K use AS
size > 64K use Noop
9 June 2005 FastOS PI Meeting 38
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Read/Write Patterns while compiling Linux source
0
1
1 26 51 76 101 126 151 176 201 226 251 276 301 326 351 376 401 426 451 476 501 526 551 576 601 626 651 676 701 726 751 776 801 826 851
Time (sec)
Read
/Wri
te
Linux Compilation Read/Write Access
9 June 2005 FastOS PI Meeting 39
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
• Performance comparison of Linux Compilation (gmake –j 16; xeon 2.78GHz; source on RAID-0 with 4 IDE drives)
Scheduler Time (Sec)
AS 830
Deadline 850
CFQ 860
Noop 848
Adaptive 843
preliminary results – inconclusive
9 June 2005 FastOS PI Meeting 40
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Lessons Learned• 80 scheduler switches• Switching is not “costing our life”• Switching has to be on a coarser granularity
– “Smaller” picture (focus: requests) is captured– “Bigger” picture (focus: workloads) needs to be
captured– Prediction does not have to be per-request– Example desired prediction: database workload
followed by streaming reads – should select deadline/cfq followed by AS
9 June 2005 FastOS PI Meeting 41
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Applications/Benchmarks used for
I/O Scheduling Research - 1 • Flexible File System Benchmark – FFSB bench
– used to generate profile-driven, I/O workload with a characteristic I/O access pattern of a given type of server, e.g., web, file, email servers, and MetaData server.
• Microbenchmarks– Streaming writes and chunk reads– Streaming reads and chunk reads– Chunk reads
9 June 2005 FastOS PI Meeting 42
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Applications/Benchmarks used for I/OScheduling Research - 2
• MADbench: I/O intensive (Borril, et al.)– based on MADCAP, an application for estimating the power
spectrum of cosmic microwave background radiation– retains computational intensity, operational complexity, and
system requirements of MADCAP and implements its three main processing steps1. builds signal correlation derivative matrices and it requires neither
read operations nor communication2. builds a subset of a data correlation matrix and then inverts it; does
not require writes3. reads a subset of the signal correlation matrices built in the first step
and performs matrix multiplication against the inverted matrix obtained in step two; does not require writes
9 June 2005 FastOS PI Meeting 43
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Process Scheduling (stalled)
Several references indicate that the Linux 2.4 scheduling policy has an adverse effect on non-real-time, non-interactive applications—a description that fits the high-performance applications
9 June 2005 FastOS PI Meeting 44
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Process Scheduling• WhatWhat to adapt: process scheduler
• Dynamic selection of “appropriate” process scheduler for observed “system state”
• WhenWhen to adapt• Change in “system state” • Below threshold related to time spent in scheduler or
length of “runnable” queue• HowHow to adapt
• Change process type to real-time, i.e., scheduler to round-robin
• A priori knowledge of process IDs• HowHow to measure effects of adaptation
• Execution time
9 June 2005 FastOS PI Meeting 45
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Too much time in scheduler
Adaptation to round-robin
Hackbench 2.4 execution times for unmodified (blue) and adapted scheduler
0
20
40
60
80
100
120
140
160
20 40 60 80 100 120 140 160 180 200 220
Groups
seco
nd
s
2.4 Baseline
2.4 Queue Lengthadaptation
2.4 Time in sched.adaptation
Adaptation: timeshared to real time (round-robin) scheduling with fixed quantum and priority per process for the lifetime of the application
9 June 2005 FastOS PI Meeting 46
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Hackbench Speedup wrt 2.4 unmodified kernel
0.000.501.001.502.002.503.00
Groups
2.4 Q. Length
2.4 Time in sched.
9 June 2005 FastOS PI Meeting 47
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Learning-based, Heuristic-driven Dynamic Adaptation (longer term)
• I/O scheduler parameter selection• IBM-Austin: genetic algorithms• UTEP: investigate potentially lower cost hybrid
techniques that combine the use of neural networks, genetic algorithms, and fuzzy logic (possible dissertation topic: Luis Ortiz, UTEP)
• Master’s thesis: neural network approach to selecting parameters of the Anticipatory Scheduler of Linux 2.6 [Moilanen]
9 June 2005 FastOS PI Meeting 48
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Near-term Research Directions – 1
• Adaptation target investigation via self-propelled instrumentation to obtain function-level traces from applications and the kernel [Mirgorodskiy and B. Miller]– Possible applications:
• LBMHD (plasma physics), PARATEC (material science), CACTUS (astrophysics), and GTC (magnetic fusion), which are able to fully utilize the performance of machines comparable to the Earth Simulator and Cray X1 [Oliker, et al.]
• Sweep3D (using for daemon control investigation)• SPECjAppServer2004 (Websphere, DB2)
9 June 2005 FastOS PI Meeting 49
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
Near-term Research Directions – 2
• Study of applications to determine profitability of – Dynamically adapting page size [Cascaval, et al.] – Daemon control (started) [D. Bailey and Hoisie, et al.]– Virtual memory management (just starting)– Dynamically adapting time quantum to control, e.g.,
process cache contention
• k-factor analysis to develop mathematical models to guide I/O parameter set selection
9 June 2005 FastOS PI Meeting 50
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
References – 1• Bailey, D., Private Communications, 2005.• Borril, J., J. Carter, L. Oliker, D. Skinner, and R. Biswas, “Integrated
Performance Monitoring of a Cosmology Application on Leading HEC Platforms,” Proceedings of the 2005 International Conference on Parallel Processing (ICPP-05), June 2005.
• Cascaval, C., E. Duesterwald, P. Sweeney, and R. Wisniewski, “Multiple Page Size Modeling and Optimization,” Proceedings of the Fourteenth International Conference on Parallel Architectures and Compilation Techniques (PACT-2005), September 2005.
• http://www.sourceforge.net/projects/ffsb
9 June 2005 FastOS PI Meeting 51
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
References – 2• Hoisie, A., D. Kerbyson, S. Pakin, F. Petrini, H. Wasserman, and J.
Fernandez-Peinador, “Identifying and Eliminating the Performance Variability on the ASCI Q Machine,” Technical Report LA-UR-03-0138, Performance and Architecture Lab, Los Alamos National Laboratory, January 2003.
• Mirgorodskiy, A., and B. Miller, “Autonomous Analysis of Interactive Systems with Self-propelled Instrumentation,” Proceedings of the Multimedia Computing and Networking Conference, 2004.
• Moilanen, J., “Genetic Algorithms in the Kernel,” http://kernel.jakem.net/, 2005.
• Oliker, L., A. Canning, J. Carter, J. Shalf, and S. Ethier, S. “Scientific Computations on Modern Parallel Vector Systems,” Proceedings of the 2004 ACM/IEEE Conference on Supercomputing, November 2004.
9 June 2005 FastOS PI Meeting 52
Dynamic Adaptability in Support of Extreme Scale University of Wisconsin-Madison
University of Texas-El Paso
References – 3• Pratt, S., and D. Heger, IBM-Austin, “Workload Dependent
Performance Evaluation of Linux 2.6 I/O schedulers,” Linux Symposium, 2, July 2004, pp. 425-448.
• Tamches, A., and B. Miller, “Dynamic Kernel I-Cache Optimization,” Proceedings of the Workshop on Binary Translation, September 2001.