high performance computing: concepts, methods & means high capacity (throughput) computing prof....

High Performance Computing: Concepts, Methods & Means

High Capacity (Throughput) Computing

Prof. Thomas SterlingDepartment of Computer Science

Louisiana State University

January 25th, 2007

Topics

• Key terms and concepts• Basic definitions• Models of parallelism• Speedup and Overhead• Capability Computing & Unix utilities• Condor : Overview• Condor : Useful commands• Performance Issues in Capacity Computing• Material for Test

2

Topics


3

Key Terms and Concepts

4

Problem

instructionsinstructions

CPU

Conventional serial execution where the problem is represented as a series of instructions that are executed by the CPU

CPU CPU CPU CPU

instructionsinstructions

Task Task Task TaskProblemProblem

Parallel execution of a problem involves partitioning of the problem into multiple executable parts that are mutually exclusive and collectively exhaustive represented as a partially ordered set exhibiting concurrency.

Parallel computing takes advantage of concurrency to :• Solve larger problems within

bounded time• Save on Wall Clock Time• Overcoming memory constraints• Utilizing non-local resources

Key Terms and Concepts

• Speedup : Relative reduction of execution time of a fixed size workload through parallel execution

• Efficiency : Ratio of the actual performance to the best possible performance.

5

processorsNontimeexecution

processoroneontimeexecutionSpeedup

____

____

)______(

____

processorsofnumberprocessorsmultipleontimeexecution

processoroneontimeexecutionEfficiency

Topics


6

Defining the 3 C’s …

• Main Classes of computing :– High capacity parallel computing : A strategy for

employing distributed computing resources to achieve high throughput processing among decoupled tasks. Aggregate performance of the total system is high if sufficient tasks are available to be carried out concurrently on all separate processing elements. No single task is accelerated.

– High capability parallel computing : A strategy for employing tightly couple structures of computing resources to achieve reduced execution time of a given application through partitioning in to concurrently executable tasks.

– Cooperative computing : A strategy for employing moderately coupled ensemble of computing resources to increase size of the data set of a user application while limiting its execution time.

7

Defining the 3 C’s …

• High capacity computing systems emphasize the overall work performed over a fixed time period. Work is defined as the aggregate amount of computation performed across all functional units, all threads, all cores, all chips, all coprocessors and network interface cards in the system.

• High capability computing systems emphasize improvement (reduction) in execution time of a single user application program of fixed data set size.

8

Adapted from : High-performance throughput computing S Chaudhry, P Caprioli, S Yip, M Tremblay - IEEE Micro, 2005 - doi.ieeecomputersociety.org

http://doi.ieeecomputersociety.org/10.1109/MM.2005.49



Topics


9

Models of Parallel Processing • Conventional models of parallel processing

– Decoupled Work Queue (covered in segment 1)– Shared memory multiple thread (covered in segment 2)– Communicating Sequential Processing (CSP message passing)

(covered in segment 3)

• Alternative models of parallel processing– SIMD

• Single instruction stream multiple data stream processor array

– Vector Machines• Hardware execution of value sequences to exploit pipelining

– Systolic• An interconnection of basic arithmetic units to match algorithm

– Data Flow• Data precedent constraint self-synchronizing fine grain execution units

supporting functional (single assignment) execution10

Decoupled Work Queue Model

• Concurrent disjoint tasks • Parametric Studies

– SPMD (single program multiple data)

• Very coarse grained• Example software package : Condor• Processor farms and clusters • Last part of Segment 1 covers this

model of parallelism

11

Shared memory multiple Thread

• Static or dynamic• Fine Grained• OpenMP• Distributed shared memory systems• Covered in Segment 2

12

Network

CPU 1 CPU 2 CPU 3

memory memory memory

Network

CPU 1 CPU 2 CPU 3


Symmetric Multi Processor (SMP usually cache coherent )

Distributed Shared Memory (DSM often not cache coherent)

Caches and Cache Coherence

• Caches are part of memory hierarchy• Match high processor demand to high capacity, long

access time main memory• Caches are low capacity (relatively) short access

time• Caches hold temporary copies of data in memory but

they can’t hold all of it at any one time• Processors may share memory but caches are

private• Cache coherence keeps copies of the same data

consistent across caches of different processors• This is tough to do, slows things down, and doesn’t

scale very well

SID

E B

AR

Dis

cuss

ion

Shared Memory Multiple Thread Model

14

• Hardware view : All system memory directly accessible from all processing elements, possibly with cache coherence. Instruction streams performed concurrently, possibly switching contexts for shared resources.

• Software view : A flow of control in a process that can have concurrent execution paths. Threads share the same address space and have self contained state information. Synchronization achieved through shared variables in memory.

• Advantages : – Threads are inexpensive to create, represent and destroy.

– Relatively faster to switch context between threads than processes.

– Better resource management due to shared address spaces utilization

• Disadvantages :– Cache coherent and symmetric multiprocessor systems not scalable

– Scalable systems are none uniform memory access (NUMA)

– System call through threads can potentially block the process and consequently degrading CPU utilization.

• More in-depth coverage on this topic in segment 2 of the course

Communicating Sequential Processes

• One process is assigned to each processor

• Work done by the processor is performed on the local data

• Data values are exchanged by messages• Synchronization constructs for inter

process coordination• Distributed Memory• Coarse Grained• MPI• Clusters and MPP

– MPP is acronym for “Massively Parallel Processor”

• Covered in Segment 3

15

Network

CPU 1 CPU 2 CPU 3


Distributed Memory (DM often not cache coherent)

Topics


16

Ideal Speedup Example

17

W

220

w1 w210 210

P28

210 210 210 210

Processors

212

P1

T(1)=220

T(28)=212

812

20

22

2Speedup

1222

2 0812

20

Efficiency

Units : steps

i

iwW

Ideal Speedup Issues

18

• W is total workload measured in elemental pieces of work (e.g. operations, instructions, etc.)

• T(p) is total execution time measured in elemental time steps (e.g. clock cycles) where p is # of execution sites (e.g. processors, threads)

• wi is work for a given task i

• Example: here we divide a million (really Mega) operation workload, W, in to a thousand tasks, w1 to w1024 each of a 1 K operations

• Assume 256 processors performing workload in parallel

• T(256) = 4096 steps, speedup = 256, Eff = 1

Granularities in ParallelismOverhead

• The additional work that needs to be performed in order to manage the parallel resources and concurrent abstract tasks that is in the critical time path.

Coarse Grained• Decompose problem into large independent

tasks. Usually there is no communication between the tasks. Also defined as a class of parallelism where: “relatively large amounts of computational work is done between communication”

Fine Grained • Decompose problem into smaller inter-

dependent tasks. Usually these tasks are usually communication intensive. Also defined as a class of parallelism where: “relatively small amounts of computational work is done between communication events” –www.llnl.gov/computing/tutorials/parallel_comp

19Images adapted from : http://www.mhpcc.edu/training/workshop/parallel_intro/

Overhead

Computation

Coarse Grained

Overhead

Computation

Finely Grained

Overhead

20

vv ww

W=4v+4wW=4v+4w

P

Wwi

wvT

PW

P

WPP

PW

W

PWW

T

TS

P

1

11

v = overheadw = work unitW = Total workTi = execution time with i processorsP = # processors

P

iiwW

1

Assumption : Workload is infinitely divisible

Overhead

21

• Overhead: Additional critical path (in time) work required to manage parallel resources and concurrent tasks that would not be necessary for purely sequential execution

• V is total overhead of workload execution

• vi is overhead for individual task wi

• Each task takes v+w time steps to complete• Overhead imposes upper bound on scalability

Scalability & Overhead

22

gg

P

gP

gg

ggP

gg

wv

P

wv

PW

W

T

TS

w

v

P

WT

w

v

P

Wvw

Pw

Wvw

P

JT

WvWT

w

W

w

WtasksJ

11

1

1)(

#

1

1 when W >> v

v = overheadwg = work unitW = Total workTi = execution time with i processorsP = # ProcessorsJ = # Tasks

Scalability and Overhead for fixed sized work tasks

23

• W is divided in to J wg sized tasks

• Each task requires v overhead work to manage• For P processors there are approximates J/P tasks

to be performed in sequence so,

• TP is J(wg + v)/P

• Note that S = T1 / TP

• So, S = P / (1 + v / wg)

Topics


24

Capacity Computing with basic Unix tools

• Combination of common Unix utilities such as ssh, scp, rsh, rcp can be used to remotely create jobs. ( to get more information about these commands try man ssh, man scp, man rsh, man rcp on any Unix shell )

• For small workloads it can be convenient to translate the execution of the program into a simple shell script.

• Relying on simple Unix utilities poses several application management constraints for cases such as :– Aborting started jobs – Querying for free machines– Querying for job status – Retrieving job results – etc..

25

Demo 1

Using Unix utilities to executing Capability Computing

26

BOINC , Seti@Home

• BOINC (Berkley Open Infrastructure for Network Computing)• Opensource software that enables distributed coarse grained

computations over the internet. • Follows the Master-Worker model, in BOINC : no

communication takes place among the worker nodes • SETI@Home• Einstein@Home• Climate prediction• And many more…

27

Topics


28

Management Middleware : Condor

• Designed, developed and maintained at University of Wisconsin Madison by a team lead by Miron Livny

• Condor is a versatile workload management system for managing pool of distributed computing resources to provide high capacity computing.

• Assists job management by providing mechanisms for job queuing, scheduling, priority management, tools that facilitate utilization of resources across Condor pools

• Condor also enables resource management by providing monitoring utilities, authentication & authorization mechanisms, condor pool management utilities and support for Grid Computing middlewares such as Globus.

• Condor Components• ClassAds• Matchmaker• Problem Solvers

29

Condor Components : Class Ads• ClassAds (Classified Advertisements) concept is

very similar to the newspaper classifieds concepts where buyers and sellers advertise their products using abstract yet uniquely defining named expressions. Example : Used Car Sales

• ClassAds language in Condor provides well defined means of describing the User Job and the end resources ( storage / computational ) so that the Condor MatchMaker can match the job with the appropriate pool of resources.


Src : Douglas Thain, Todd Tannenbaum, and Miron Livny, "Distributed Computing in Practice: The Condor Experience" Concurrency and Computation: Practice and

Experience, Vol. 17, No. 2-4, pages 323-356, February-April, 2005.http://www.cs.wisc.edu/condor/doc/condor-practice.pdf 30

Job ClassAd & Machine ClassAd

Condor MatchMaker• MatchMaker, a crucial part of the Condor

architecture, uses the job description classAd provided by the user and matches the Job to the best resource based on the Machine description classAd

• MatchMaking in Condor is performed in 4 steps : 1. Job Agent (A) and resources (R) advertise themselves.

2. Matchmaker (M) processes the known classAds and generates pairs that best match resources and jobs

3. Matchmaker informs each party of the job-resource pair of their prospective match.

4. The Job agent and resource establish connection for further processing. (Matchmaker plays no role in this step, thus ensuring separation between selection of resources and subsequent activities)


Src : Douglas Thain, Todd Tannenbaum, and Miron Livny, "Distributed Computing in Practice: The Condor Experience" Concurrency and

Computation: Practice and Experience, Vol. 17, No. 2-4, pages 323-356, February-April, 2005.

http://www.cs.wisc.edu/condor/doc/condor-practice.pdf

32

Condor Problem Solvers• Master-Worker (MW) is a problem solving system that is

useful for solving a coarse grained problem of indeterminate size such as parameter sweep etc.

• The MW Solver in Condor consists of 3 main components : work-list, a tracking module, and a steering module. The work-list keeps track of all pending work that master needs done. The tracking module monitors progress of work currently in progress on the worker nodes. The steering module directs computation based on results gathered and the pending work-list and communicates with the matchmaker to obtain additional worker processes.

• DAGMan is used to execute multiple jobs that have dependencies represented as a Directed Acyclic Graph where the nodes correspond to the jobs and edges correspond to the dependencies between the jobs. DAGMan provides various functionalities for job monitoring and fault tolerance via creation of rescue DAGs.


33

MasterMaster

w1w1 w..Nw..N

Core components of Condor• condor_master: This program runs constantly and ensures that all other parts of Condor

are running. If they hang or crash, it restarts them. • condor_collector: This program is part of the Condor central manager. It collects

information about all computers in the pool as well as which users want to run jobs. It is what normally responds to the condor_status command. It's not running on your computer, but on the main Condor pool host (Celeritas head node).

• condor_negotiator: This program is part of the Condor central manager. It decides what jobs should be run where. It's not running on your computer, but on on the main Condor pool host (Celeritas head node).

• condor_startd: If this program is running, it allows jobs to be started up on this computer--that is, hal is an "execute machine". This advertises hal to the central manager (more on that later) so that it knows about this computer. It will start up the jobs that run.

• condor_schedd If this program is running, it allows jobs to be submitted from this computer--that is, hal is a "submit machine". This will advertise jobs to the central manager so that it knows about them. It will contact a condor_startd on other execute machines for each job that needs to be started.

• condor_shadow For each job that has been submitted from this computer, there is one condor_shadow running. It will watch over the job as it runs remotely. In some cases it will provide some assistance You may or may not see any condor_shadow processes running, depending on what is happening on the computer when you try it out.

35Source : http://www.cs.wisc.edu/condor/tutorials/cw2005-condor/intro.html

Topics


36

Condor : A Walkthrough of Condor commands

condor_status : provides current pool status

condor_q : provides current job queue

condor_submit : submit a job to condor pool

condor_rm : delete a job from job queue

37

What machines are available ? (condor_status)

condor_status queries resource information sources and provides the current status of the condor pool of resources

38

Some common condor_status command line options : -help : displays usage information -avail : queries condor_startd ads and prints information about available

resources -claimed : queries condor_startd ads and prints information about

claimed resources -ckptsrvr : queries condor_ckpt_server ads and display checkpoint

server attributes -pool hostname queries the specified central manager (by default

queries $COLLECTOR_HOST) -verbose : displays entire classads For more options and what they do run “condor_status –help”

condor_status : Resource States

• Owner : The machine is currently being utilized by a user. The machine is currently unavailable for jobs submitted by condor until the current user job completes.

• Claimed : Condor has selected the machine for use by other users.

• Unclaimed : Machine is unused and is available for selection by condor.

• Matched : Machine is in a transition state between unclaimed and claimed

• Preempting : Machine is currently vacating the resource to make it available to condor.

39

Example : condor_status

40

[cdekate@celeritas ~]$ condor_status

Name OpSys Arch State Activity LoadAv Mem ActvtyTime

vm1@compute-0 LINUX X86_64 Unclaimed Idle 0.000 1964 3+13:42:23vm2@compute-0 LINUX X86_64 Unclaimed Idle 0.000 1964 3+13:42:24vm3@compute-0 LINUX X86_64 Unclaimed Idle 0.010 1964 0+00:45:06vm4@compute-0 LINUX X86_64 Owner Idle 1.000 1964 0+00:00:07vm1@compute-0 LINUX X86_64 Unclaimed Idle 0.000 1964 3+13:42:25vm2@compute-0 LINUX X86_64 Unclaimed Idle 0.000 1964 1+09:05:58vm3@compute-0 LINUX X86_64 Unclaimed Idle 0.000 1964 3+13:37:27vm4@compute-0 LINUX X86_64 Unclaimed Idle 0.000 1964 0+00:05:07……vm3@compute-0 LINUX X86_64 Unclaimed Idle 0.000 1964 3+13:42:33vm4@compute-0 LINUX X86_64 Unclaimed Idle 0.000 1964 3+13:42:34

Total Owner Claimed Unclaimed Matched Preempting Backfill

X86_64/LINUX 32 3 0 29 0 0 0

Total 32 3 0 29 0 0 0

What jobs are currently in the queue? condor_q

• condor_q provides a list of job that have been submitted to the Condor pool

• Provides details about jobs including which cluster the job is running on, owner of the job, memory consumption, the name of the executable being processed, current state of the job, when the job was submitted and how long has the job been running.

41

Some common condor_q command line options : -global : queries all job queues in the pool -name : queries based on the schedd name provides a queue listing of the

named schedd -claimed : queries condor_startd ads and prints information about claimed

resources -goodput : displays job goodput statistics (“goodput is the allocation time

when an application uses a remote workstation to make forward progress.” – Condor Manual)

-cputime : displays the remote CPU time accumulated by the job to date... For more options run : “condor_q –help”

[cdekate@celeritas ~]$ condor_q

-- Submitter: celeritas.cct.lsu.edu : <130.39.128.68:40472> : celeritas.cct.lsu.edu ID OWNER SUBMITTED RUN_TIME ST PRI SIZE CMD 30.0 cdekate 1/23 07:52 0+00:01:13 R 0 9.8 fib 100 30.1 cdekate 1/23 07:52 0+00:01:09 R 0 9.8 fib 100 30.2 cdekate 1/23 07:52 0+00:01:07 R 0 9.8 fib 100 30.3 cdekate 1/23 07:52 0+00:01:11 R 0 9.8 fib 100 30.4 cdekate 1/23 07:52 0+00:01:05 R 0 9.8 fib 100

5 jobs; 0 idle, 5 running, 0 held[cdekate@celeritas ~]$

42

Example : condor_q

How to submit your Job ? condor_submit

• Create a job classAd (condor submit file) that contains Condor keywords and user configured values for the keywords.

• Submit the job classAd using “condor_submit” • Example :

condor_submit matrix.submit

• condor_submit –h provides additional flags

43

[cdekate@celeritas NPB3.2-MPI]$ condor_submit -hUsage: condor_submit [options] [cmdfile] Valid options: -verbose verbose output -name <name> submit to the specified schedd -remote <name> submit to the specified remote schedd (implies -spool) -append <line> add line to submit file before processing (overrides submit file; multiple -a lines ok) -disable disable file permission checks -spool spool all files to the schedd -password <password> specify password to MyProxy server -pool <host> Use host as the central manager to query If [cmdfile] is omitted, input is read from stdin

condor_submit : Example

44

[cdekate@celeritas ~]$ condor_submit fib.submit Submitting job(s).....Logging submit event(s).....5 job(s) submitted to cluster 35.[cdekate@celeritas ~]$ condor_q

-- Submitter: celeritas.cct.lsu.edu : <130.39.128.68:51675> : celeritas.cct.lsu.edu ID OWNER SUBMITTED RUN_TIME ST PRI SIZE CMD 35.0 cdekate 1/24 15:06 0+00:00:00 I 0 9.8 fib 10 35.1 cdekate 1/24 15:06 0+00:00:00 I 0 9.8 fib 15 35.2 cdekate 1/24 15:06 0+00:00:00 I 0 9.8 fib 20 35.3 cdekate 1/24 15:06 0+00:00:00 I 0 9.8 fib 25 35.4 cdekate 1/24 15:06 0+00:00:00 I 0 9.8 fib 30

5 jobs; 5 idle, 0 running, 0 held[cdekate@celeritas ~]$

How to delete a submitted job ? condor_rm

• condor_rm : Deletes one or more jobs from Condor job pool. If a particular Condor pool is specified as one of the arguments then the condor_schedd matching the specification is contacted for job deletion, else the local condor_schedd is contacted.

45

[cdekate@celeritas ~]$ condor_rm -h Usage: condor_rm [options] [constraints] where [options] is zero or more of: -help Display this message and exit -version Display version information and exit -name schedd_name Connect to the given schedd -pool hostname Use the given central manager to find daemons -addr <ip:port> Connect directly to the given "sinful string" -reason reason Use the given RemoveReason -forcex Force the immediate local removal of jobs in the X state (only affects jobs already being removed) and where [constraints] is one or more of: cluster.proc Remove the given job cluster Remove the given cluster of jobs user Remove all jobs owned by user -constraint expr Remove all jobs matching the boolean expression -all Remove all jobs (cannot be used with other constraints)[cdekate@celeritas ~]$

[cdekate@celeritas ~]$ condor_q-- Submitter: celeritas.cct.lsu.edu : <130.39.128.68:51675> :

celeritas.cct.lsu.edu ID OWNER SUBMITTED RUN_TIME ST PRI SIZE CMD 41.0 cdekate 1/24 15:43 0+00:00:03 R 0 9.8 fib 100 41.1 cdekate 1/24 15:43 0+00:00:01 R 0 9.8 fib 150 41.2 cdekate 1/24 15:43 0+00:00:00 R 0 9.8 fib 200 41.3 cdekate 1/24 15:43 0+00:00:00 R 0 9.8 fib 250 41.4 cdekate 1/24 15:43 0+00:00:00 R 0 9.8 fib 300

5 jobs; 0 idle, 5 running, 0 held[cdekate@celeritas ~]$ condor_rm 41.4Job 41.4 marked for removal[cdekate@celeritas ~]$ condor_rm 41 Cluster 41 has been marked for removal.[cdekate@celeritas ~]$

46

condor_rm : Example

Creating Condor submit file ( Job a ClassAd )

• Condor submit file contains key-value pairs that help describe the application to condor.

• Condor submit files are job ClassAds. • Some of the common descriptions found in the job

ClassAds are :

47

executable = (path to the executable to run on Condor)input = (standard input provided as a file)output = (standard output stored in a filelog = (output to log file )arguments = (arguments to be supplied to the )queue

DEMO 2 : Steps involved in running a job on Condor.

1. Creating a Condor submit file

2. Submitting the Condor submit file to a Condor pool

3. Checking the current state of a submitted job

4. Job status Notification

48

Condor Usage Statistics

49

Montage workload implemented and executed using Condor ( Source : Dr. Dan Katz )

• Mosaicking astronomical images : • Powerful Telescopes taking high resolution (and highest zoom) pictures of the sky can cover small region over time• Problem being solved in this project is “stitching” these images together to make a high-resolution zoomed in snapshot of the sky.• Aggregate requirements of 140000 CPU hours (~16 years on a single machine) output ranging in the order of 6 TeraBytes

50

Example DAG for 10 input files

mAdd

mBackground

mBgModel

mProject

mDiff

mFitPlane

mConcatFit

Data Stage-in nodes

Montage compute nodes

Data stage-out nodes

Registration nodes

Pegasus

Grid Information Systems

Information about available resources,

data location

Grid

Condor DAGMan

Maps an abstract workflow to an executable form

Executes the workflow

MyProxy

User’s grid credentials

http://pegasus.isi.edu/

Montage Use By IPHAS: The INT/WFC Photometric H-alpha Survey of the Northern Galactic Plane (Source : Dr. Dan Katz)

Supernova remnant S147

Nebulosity in vicinity of HII region, IC 1396B, in Cepheus

Crescent Nebula NGC 6888

Study extreme phases of stellar

evolution that involve very large

mass loss

Topics


52

53

• Throughput computing• Performance measured as total workload performed over time

to complete• Overhead factors

– Start up time

– Input data distribution

– Output result data collection

– Terminate time

– No task coupling or inter-task coordination overhead

• Starvation– Insufficient work to keep all processors busy

– Inadequate parallelism of coarse grained task parallelism

– Poor or uneven load distribution

Capacity Computing Performance Issues

Topics


54

Summary : Material for the Test• Understand material on slide (4,5),(7,8)• Understand example detailed in slides 17, 18• Understand (19) and be able to derive (20,21), (22,

23)• Understand Condor concepts detailed in slides

30,31,32• Condor Commands (37-47) : know what the basic

commands are, what they do and interpret output presented by them etc. (No need to memorize command-line options)

• Understand issues listed on slide 53• Required reading materials :

– http://www.cct.lsu.edu/~cdekate/7600/beowulf-chapter-rev1.pdf

– Specific pages to focus on : 3-16 55

http://www.cct.lsu.edu/~cdekate/7600/beowulf-chapter-rev1.pdf

high performance computing: concepts, methods & means high capacity (throughput) computing prof....

Documents

high performance computing

cooperative computing

capacity computingmaterial

cs main classes of computing

reduced execution time

high throughput processing

orgtopicskey terms

fixed time period