Version: 0.6-5
Date: 2022-06-25
Title: High Performance Cluster Models Based on Kiefer-Wolfowitz Recursion
Author: Alexander Rumyantsev [aut, cre]
Maintainer: Alexander Rumyantsev <ar0@sampo.ru>
Description: Probabilistic models describing the behavior of workload and queue on a High Performance Cluster and computing GRID under FIFO service discipline basing on modified Kiefer-Wolfowitz recursion. Also sample data for inter-arrival times, service times, number of cores per task and waiting times of HPC of Karelian Research Centre are included, measurements took place from 06/03/2009 to 02/30/2011. Functions provided to import/export workload traces in Standard Workload Format (swf). Stability condition of the model may be verified either exactly, or approximately. Stability analysis: see Rumyantsev and Morozov (2017) <doi:10.1007/s10479-015-1917-2>, workload recursion: see Rumyantsev (2014) <doi:10.1109/PDCAT.2014.36>.
License: GPL-2 | GPL-3 [expanded from: GPL (≥ 2)]
RoxygenNote: 7.1.2
Repository: CRAN
Repository/R-Forge/Project: hpcwld
Repository/R-Forge/Revision: 27
Repository/R-Forge/DateTimeStamp: 2022-06-25 20:09:01
Date/Publication: 2022-06-26 21:20:07 UTC
NeedsCompilation: no
Packaged: 2022-06-25 20:31:06 UTC; rforge
Depends: R (≥ 2.10)

Model and data for High Performance Cluster workload

Description

This package contains several models describing the behavior of workload and queue on a High Performance Cluster and computing GRID under FIFO service discipline basing on modified Kiefer-Wolfowitz recursion. Also sample data for inter-arrival times, service times, number of cores per task and waiting times of HPC of Karelian Research Centre are included, measurements took place from 06/03/2009 to 02/30/2011. The stability condition of the model can be verified either exactly, or approximately.

Details

Package: hpcwld
Type: Package
Version: 0.5
Date: 2015-02-14
License: GNU GPL
LazyLoad: yes

Author(s)

Alexander Rumyantsev (Institute of Applied Mathematical Research, Karelian Research Centre, RAS)

References

E.V. Morozov, A.Rumyantsev. Stability analysis of a multiprocessor model describing a high performance cluster. XXIX International Seminar on Stability Problems for Stochastic Models and V International Workshop "Applied Problems in Theory of Probabilities and Mathematical Statistics related to modeling of information systems". Book of Abstracts. 2011. Pp. 82–83.

A. Rumyantsev. Simulating Supercomputer Workload with hpcwld package for R // Proceedings of 2014 15th International Conference on Parallel and Distributed Computing, Applications and Technologies. IEEE, 2014. P. 138-143. URL: http://conferences.computer.org/pdcat/2014/papers/8334a138.pdf

A. Rumyantsev. Evaluating the stability of supercomputer workload model // Journal on Selected Topics in Nano Electronics and Computing, Vol. 2, No. 2, December 2014. P. 36-39.

http://cluster.krc.karelia.ru

Examples

Wld(T=rexp(1000,1), S=rexp(1000,1), round(runif(1000,1,10)), 10)
# returns the workload, delay and total cpus used 
# for a cluster with 10 CPUs and random exponential times

Approximate, dynamic iterative computation of the stability constant for a workload of a High Performance Cluster model

Description

This function calculates the constant C that is used in the stability criterion of a supercomputer model, which is basically the following: lambda/mu<C, where lambda is the task arrivals rate, and mu is the service intensity. The constant depends only on the number of servers in the model and the distribution of classes of customers, where class is the number of servers required. This method of calculation allows to stop on some depth of dynamics, thus allowing to calculate an approximate value in faster time. The constant is valid only for the model with simultaneous service.

Usage

ApproxC(s, p, depth = 3)

Arguments

s

number of servers in the model

p

vector of class distribution

depth

By default calculates up to groups of 3 tasks. When depth=s, calculates the exact value. However, depth=s might take a bit more time.

Value

The value of a constant C in the relation lambda/mu < C is returned

Examples

ApproxC(s=2,p=c(.5,.5), depth=3)

Distributional Measure of Correlation

Description

This is a suggested by Dror Feitelson measure of correlation for dependent variables, that may be successfully used to examine the datasets from a High Performance Cluster logs

Usage

DMC(X, Y)

Arguments

X

First variable (vector)

Y

Second variable (vector)

Value

One value between -1 and 1, characterizing the dependence between the variables

References

http://interstat.statjournals.net/YEAR/2004/abstracts/0412001.php?Name=412001

Examples

data(HPC_KRC)
DMC(HPC_KRC$service[1:1000], HPC_KRC$cores_requested[1:1000])

Convertor from a dataframe to Standart Workload Format

Description

Note that this is only a wrapper for the ToSWF command with a dataframe argument. It needs a correctly built dataframe and converts it to a Standart Workload Format used to share the logfiles of High Performance Clusters

Usage

DataToSWF(Frame, filename = "output.swf")

Arguments

Frame

A dataframe containing the variables needed by ToSWF function

filename

The file to store the converted workload (output.swf by default)

Details

The Standart Workload Format is a single format to store and exchange high performance cluster logs, that is used in Parallel Workload Archive. See references for current standard. The SWF format may contain additional data, but in this package only the 1st to 5th fields are used. One may also need to manually fill in the header of the file in order to completely prepare the resulting SWF file.

Value

Nothing is returned, but a file is created in the current working directory (with default name output.swf) containing the converted data.

References

Feitelson, D.G. and Tsafrir, D. and Krakov D. 2012 Experience with the Parallel Workloads Archive. Technical Report 2012-6, School of Computer Science and Engineering, the Hebrew University April, 2012, Jerusalem, Israel

https://www.cs.huji.ac.il/labs/parallel/workload/swf.html

Examples

## Not run: 
data(HPC_KRC)
tmp=data.frame(T=HPC_KRC$interarrival, S=HPC_KRC$service, N=HPC_KRC$cores_used, D=HPC_KRC$delay)
DataToSWF(tmp)

## End(Not run)

Convertor to a dataset from a Standart Workload Format

Description

This is a convertor from a Standart Workload Format (used to share the logfiles of High Performance Clusters) to an internally used in a package dataset format

Usage

FromSWF(filename)

Arguments

filename

A mandatory field containing the path to SWF file

Details

The Standart Workload Format is a single format to store and exchange high performance cluster logs, that is used in Parallel Workload Archive. See references for current standard. The SWF format may contain additional data, but in this package only the 1st to 5th fields are used. One may also need to manually fill in the header of the file in order to completely prepare the resulting SWF file.

Value

A dataset is returned, containing 'delay' as a vector of delays exhibited by each task, 'total_cores' as the total busy CPUs in time of arrival of each task, and 'workload' as total work left at each CPU.

References

Feitelson, D.G. and Tsafrir, D. and Krakov D. 2012 Experience with the Parallel Workloads Archive. Technical Report 2012-6, School of Computer Science and Engineering, the Hebrew University April, 2012, Jerusalem, Israel

https://www.cs.huji.ac.il/labs/parallel/workload/swf.html

Workload data for High Performance Cluster of High Performance Data Center of Karelian Research Center, Russian Academy of Sciences.

Description

This is a complete data of the tasks which successfully finished executions at HPC of HPDC KRC RAS for time period 06/03/2009 to 02/04/2011, a total of 8282 tasks. The data contains interarrival times, service times, cores that tasks requested, cores really used (due to administrative limitations) and delays excursed by tasks, all in seconds.

Usage

data(HPC_KRC)

Format

A data frame with 8281 observations on the following 5 variables.

interarrival

a numeric vector

service

a numeric vector

cores_requested

a numeric vector

cores_used

a numeric vector

delays

a numeric vector

Source

http://cluster.krc.karelia.ru

References

http://cluster.krc.karelia.ru

Examples

data(HPC_KRC)

Workload data for High Performance Cluster of High Performance Data Center of Karelian Research Center, Russian Academy of Sciences.

Description

This is a complete data of the tasks which successfully finished executions at HPC of HPDC KRC RAS for time period 02/04/2011 to 16/04/2012, a total of 9389 tasks. The data contains interarrival times, service times, cores that tasks used, and delays excursed by tasks, all in seconds.

Usage

data(HPC_KRC2)

Format

A data frame with 9389 observations on the following 3 variables.

interarrival

a numeric vector

service

a numeric vector

cores_used

a numeric vector

delays

a numeric vector

Source

http://cluster.krc.karelia.ru

References

http://cluster.krc.karelia.ru

Examples

data(HPC_KRC2)

This function gives the maximal throughput of a two-server supercomputer (Markov) model with various service speeds, various rates of classes and random speed scaling at arrival/depature

Description

This function gives the maximal throughput of a two-server supercomputer (Markov) model with various service speeds, various rates of classes and random speed scaling at arrival/depature

Usage

MaxThroughput2(p1, pa, pd, mu1, mu2, f1, f2)

Arguments

p1

probability of class 1 arrival

pa

probability of speed switch from f1 to f2 upon arrival

pd

probability of speed switch from f2 to f1 upon departure

mu1

work amount parameter (for exponential distribution) for class 1

mu2

work amount parameter (for exponential distribution) for class 2

f1

low speed (workunits per unit time)

f2

high speed (workunits per unit time)

Value

maximal input rate, that is the stability boundary

Convertor from a dataset to Standart Workload Format

Description

This is a convertor from a correctly built dataset to a Standart Workload Format used to share the logfiles of High Performance Clusters

Usage

ToSWF(T, S, N, D, filename = "output.swf")

Arguments

T

Interarrival times of tasks (a vector)

S

Service times of tasks (a vector)

N

Number of cores each task needs (a vector)

D

The delays of tasks in a queue (a vector)

filename

The file to store the converted workload (output.swf by default)

Details

The Standart Workload Format is a single format to store and exchange high performance cluster logs, that is used in Parallel Workload Archive. See references for current standard. The SWF format may contain additional data, but in this package only the 1st to 5th fields are used. One may also need to manually fill in the header of the file in order to completely prepare the resulting SWF file.

Value

Nothing is returned, but a file is created in the current working directory (with default name output.swf) containing the converted data.

References

Feitelson, D.G. and Tsafrir, D. and Krakov D. 2012 Experience with the Parallel Workloads Archive. Technical Report 2012-6, School of Computer Science and Engineering, the Hebrew University April, 2012, Jerusalem, Israel

https://www.cs.huji.ac.il/labs/parallel/workload/swf.html

Examples

## Not run: 
data(HPC_KRC)
ToSWF(HPC_KRC$interarrival, HPC_KRC$service, HPC_KRC$cores_requested, HPC_KRC$delay)

## End(Not run)

Workload of a High Performance Cluster model

Description

This function computes the Kiefer-Wolfowitz modified vector for a HPC model. This vector contains the work left on each of 'm' servers of a cluster for the time of the arival of a task. Two methods are available, one for the case of concurrent server release (all the servers end a single task simultaneously), other for independent release (service times on each server are independent).

Usage

Wld(T, S, N, m, method = "concurrent")

Arguments

T

Interarrival times of tasks

S

Service times of customers (a vector of length n, or a matrix nrows=n, ncols='m').

N

Number of servers each customer needs

m

Number of servers for a supercomputer

method

Independent or concurrent

Value

A dataset is returned, containing 'delay' as a vector of delays exhibited by each task, 'total_cores' as the total busy CPUs in time of arrival of each task, and 'workload' as total work left at each CPU.

Examples

Wld(T=rexp(1000,1), S=rexp(1000,1), round(runif(1000,1,10)), 10)

Dataset with raw workload data from HPDC KRC RAS

Description

Source data for workload of HPC of HPDC KRC RAS. More usable dataset is HPC_KRC. This are raw times in sec. since 1 January 1970, for tasks arrival times, start of execution times and end times.

Usage

data(X)

Format

The format is: num [1:8499, 1:3] 1.24e+09 1.24e+09 1.24e+09 1.24e+09 1.24e+09 ...

Source

http://cluster.krc.karelia.ru

References

http://cluster.krc.karelia.ru

Examples

data(X)