workload modeling and its effect on performance evaluation dror feitelson hebrew university

149
Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Post on 21-Dec-2015

220 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Workload Modelingand its Effect on

Performance Evaluation

Dror Feitelson

Hebrew University

Page 2: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Performance Evaluation

• In system design– Selection of algorithms– Setting parameter values

• In procurement decisions– Value for money– Meet usage goals

• For capacity planing

Page 3: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

The Good Old Days…

• The skies were blue

• The simulation results were conclusive

• Our scheme was better than theirs

Feitelson & Jette, JSSPP 1997

Page 4: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

But in their papers,

Their scheme was better than ours!

Page 5: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

How could they be so wrong?

Page 6: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

• The system’s design(What we teach in algorithms and data structures)

• Its implementation(What we teach in programming courses)

• The workload to which it is subjected

• The metric used in the evaluation

• Interactions between these factors

Performance evaluation depends on:

Page 7: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

• The system’s design(What we teach in algorithms and data structures)

• Its implementation(What we teach in programming courses)

• The workload to which it is subjected

• The metric used in the evaluation

• Interactions between these factors

Performance evaluation depends on:

Page 8: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Outline for Today

• Three examples of how workloads affect performance evaluation

• Workload modeling– Getting data– Fitting, correlations, stationarity…– Heavy tails, self similarity…

• Research agenda

In the context of parallel job scheduling

Page 9: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Example #1

Gang Scheduling and

Job Size Distribution

Page 10: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Gang What?!?

Time slicing parallel jobs with coordinated context switching

Ousterhoutmatrix

Ousterhout, ICDCS 1982

Page 11: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Gang What?!?

Time slicing parallel jobs with coordinated context switching

Ousterhoutmatrix

Optimization:Alternativescheduling

Ousterhout, ICDCS 1982

Page 12: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Packing Jobs

Use a buddy system for allocating processors

Feitelson & Rudolph, Computer 1990

Page 13: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Packing Jobs

Use a buddy system for allocating processors

Page 14: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Packing Jobs

Use a buddy system for allocating processors

Page 15: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Packing Jobs

Use a buddy system for allocating processors

Page 16: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Packing Jobs

Use a buddy system for allocating processors

Page 17: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

The Question:

• The buddy system leads to internal fragmentation

• But it also improves the chances of alternative scheduling, because processors are allocated in predefined groups

Which effect dominates the other?

Page 18: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

The Answer (part 1):

Feitelson & Rudolph, JPDC 1996

Page 19: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

The Answer (part 2):

Page 20: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

The Answer (part 2):

Page 21: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

The Answer (part 2):

Page 22: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

The Answer (part 2):

• Many small jobs

• Many sequential jobs

• Many power of two jobs

• Practically no jobs use full machine

Conclusion: buddy system should work well

Page 23: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Verification

Feitelson, JSSPP 1996

Page 24: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Example #2

Parallel Job Scheduling

and Job Scaling

Page 25: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Variable Partitioning

• Each job gets a dedicated partition for the duration of its execution

• Resembles 2D bin packing

• Packing large jobs first should lead to better performance

• But what about correlation of size and runtime?

Page 26: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Scaling Models

• Constant work– Parallelism for speedup: Amdahl’s Law– Large first SJF

• Constant time– Size and runtime are uncorrelated

• Memory bound– Large first LJF– Full-size jobs lead to blockout

Worley, SIAM JSSC 1990

Page 27: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

“Scan” Algorithm

• Keep jobs in separate queues according to size (sizes are powers of 2)

• Serve the queues Round Robin, scheduling all jobs from each queue (they pack perfectly)

• Assuming constant work model, large jobs only block the machine for a short time

• But the memory bound model would lead to excessive queueing of small jobs

Krueger et al., IEEE TPDS 1994

Page 28: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

The Data

Page 29: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

The Data

Page 30: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

The Data

Page 31: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

The Data

Data: SDSC Paragon, 1995/6

Page 32: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

The Data

Data: SDSC Paragon, 1995/6

Page 33: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

The Data

Data: SDSC Paragon, 1995/6

Page 34: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Conclusion

• Parallelism used for better results, not for faster results

• Constant work model is unrealistic

• Memory bound model is reasonable

• Scan algorithm will probably not perform well in practice

Page 35: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Example #3

Backfilling and

User Runtime Estimation

Page 36: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Backfilling

• Variable partitioning can suffer from external fragmentation

• Backfilling optimization: move jobs forward to fill in holes in the schedule

• Requires knowledge of expected job runtimes

Page 37: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Variants

•EASY backfilling

Make reservation for first queued job

•Conservative backfilling

Make reservation for all queued jobs

Page 38: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

User Runtime Estimates

• Lower estimates improve chance of backfilling and better response time

• Too low estimates run the risk of having the job killed

• So estimates should be accurate, right?

Page 39: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

They Aren’t

Mu’alem & Feitelson, IEEE TPDS 2001

Page 40: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Surprising Consequences

• Inaccurate estimates actually lead to improved performance

• Performance evaluation results may depend on the accuracy of runtime estimates– Example: EASY vs. conservative– Using different workloads– And different metrics

Page 41: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

EASY vs. Conservative

Using CTC SP2 workload

Page 42: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

EASY vs. Conservative

Using Jann workload model

Page 43: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

EASY vs. Conservative

Using Feitelson workload model

Page 44: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Conflicting Results Explained• Jann uses accurate runtime estimates

• This leads to a tighter schedule

• EASY is not affected too much

• Conservative manages less backfilling of long jobs, because respects more reservations

Page 45: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Conservative is bad for the long jobsGood for short ones that are respected

Conservative

EASY

Page 46: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Conflicting Results Explained

• Response time sensitive to long jobs, which favor EASY

• Slowdown sensitive to short jobs, which favor conservative

• All this does not happen at CTC, because estimates are so loose that backfill can occur even under conservative

Page 47: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Verification

Run CTC workload with accurate estimates

Page 48: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

But What About My Model?

Simply does not have such small long jobs

Page 49: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Workload Data Sources

Page 50: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

No Data

• Innovative unprecedented systems– Wireless– Hand-held

• Use an educated guess– Self similarity– Heavy tails– Zipf distribution

Page 51: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Serendipitous Data

• Data may be collected for various reasons– Accounting logs– Audit logs– Debugging logs– Just-so logs

• Can lead to wealth of information

Page 52: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

NASA Ames iPSC/860 log42050 jobs from Oct-Dec 1993 user job nodes runtime date time

user4 cmd8 32 70 11/10/93 10:13:17

user4 cmd8 32 70 11/10/93 10:19:30

user42 nqs450 32 3300 11/10/93 10:22:07

user41 cmd342 4 54 11/10/93 10:22:37

sysadmin pwd 1 6 11/10/93 10:22:42

user4 cmd8 32 60 11/10/93 10:25:42

sysadmin pwd 1 3 11/10/93 10:30:43

user41 cmd342 4 126 11/10/93 10:31:32 Feitelson & Nitzberg, JSSPP 1995

Page 53: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Distribution of Job Sizes

Page 54: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Distribution of Job Sizes

Page 55: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Distribution of Resource Use

Page 56: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Distribution of Resource Use

Page 57: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Degree of Multiprogramming

Page 58: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

System Utilization

Page 59: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Job Arrivals

Page 60: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Arriving Job Sizes

Page 61: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Distribution of Interarrival Times

Page 62: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Distribution of Runtimes

Page 63: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

User Activity

Page 64: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Repeated Execution

Page 65: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Application Moldability

Page 66: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Distribution of Run Lengths

Page 67: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Predictability in Repeated Runs

Page 68: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Recurring Findings

• Many small and serial jobs

• Many power-of-two jobs

• Weak correlation of job size and duration

• Job runtimes are bounded but have CV>1

• Inaccurate user runtime estimates

• Non-stationary arrivals (daily/weekly cycle)

• Power-law user activity, run lengths

Page 69: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Instrumentation

• Passive: snoop without interfering

• Active: modify the system– Collecting the data interferes with system

behavior– Saving or downloading the data causes

additional interference– Partial solution: model the interference

Page 70: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Data Sanitation

• Strange things happen

• Leaving them in is “safe” and “faithful” to the real data

• But it risks situations in which a non-representative situation dominates the evaluation results

Page 71: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Arrivals to SDSC SP2

Page 72: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Arrivals to LANL CM-5

Page 73: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Arrivals to CTC SP2

Page 74: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Arrivals to SDSC Paragon

What are they doing at 3:30

AM?

Page 75: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

3:30 AM

• Nearly every day, a set of 16 jobs are run by the same user

• Most probably the same set, as they typically have a similar pattern of runtimes

• Most probably these are administrative jobs that are executed automatically

Page 76: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Arrivals to CTC SP2

Page 77: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Arrivals to SDSC SP2

Page 78: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Arrivals to LANL CM-5

Page 79: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Arrivals to SDSC Paragon

Page 80: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Are These Outliers?

• These large activity outbreaks are easily distinguished from normal activity

• They last for several days to a few weeks

• They appear at intervals of several months to more than a year

• They are each caused by a single user!– Therefore easy to remove

Page 81: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University
Page 82: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Two Aspects

• In workload modeling, should you include this in the model?– In a general model, probably not– Conduct separate evaluation for special

conditions (e.g. DOS attack)

• In evaluations using raw workload data, there is a danger of bias due to unknown special circumstances

Page 83: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Automation

• The idea:– Cluster daily data in based on various

workload attributes– Remove days that appear alone in a cluster– Repeat

• The problem:– Strange behavior often spans multiple days

n

Cirne &Berman, Wkshp Workload Charact. 2001

Page 84: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Workload Modeling

Page 85: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Statistical Modeling

• Identify attributes of the workload

• Create empirical distribution of each attribute

• Fit empirical distribution to create model

• Synthetic workload is created by sampling from the model distributions

Page 86: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Fitting by Moments

• Calculate model parameters to fit moments of empirical data

• Problem: does not fit the shape of the distribution

Page 87: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Jann et al, JSSPP 1997

Page 88: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Fitting by Moments

• Calculate model parameters to fit moments of empirical data

• Problem: does not fit the shape of the distribution

• Problem: very sensitive to extreme data values

Page 89: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Effect of Extreme Runtime Values

Change when top records omitted

omit mean CV

0.01% -2.1% -29%

0.02% -3.0% -35%

0.04% -3.7% -39%

0.08% -4.6% -39%

0.16% -5.7% -42%

0.31% -7.1% -42%Downey & Feitelson, PER 1999

Page 90: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Alternative: Fit to Shape

• Maximum likelihood: what distribution parameters were most likely to lead to the given observations– Needs initial guess of functional form

• Phase type distributions– Construct the desired shape

• Goodness of fit– Kolmogorov-Smirnov: difference in CDFs– Anderson-Darling: added emphasis on tail– May need to sample observations

Page 91: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Correlations

• Correlation can be measured by the correlation coefficient

• It can be modeled by a joint distribution function

• Both may not be very useful

Page 92: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University
Page 93: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Correlation Coefficient

system CC

CTC SP2 -0.029

KTH SP2 0.011

SDSC SP2 0.145

LANL CM-5 0.211

SDSCParagon 0.305

Gives low results for correlation of runtime and size in parallel systems

22yyxx

yyxx

ii

ii

Page 94: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Distributions

A restricted version of a joint distribution

Page 95: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Modeling Correlation

• Divide range of one attribute into sub-ranges

• Create a separate model of other attribute for each sub-range

• Models can be independent, or model parameter can depend on sub-range

Page 96: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Stationarity

• Problem of daily/weekly activity cycle– Not important if unit of activity is very small

(network packet)– Very meaningful if unit of work is long

(parallel job)

Page 97: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

How to Modify the Load

• Multiply interarrivals or runtimes by a factor– Changes the effective length of the day

• Multiply machine size by a factor– Modifies packing properties

• Add users

Page 98: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Stationarity

• Problem of daily/weekly activity cycle– Not important if unit of activity is very small

(network packet)– Very meaningful if unit of work is long

(parallel job)

• Problem of new/old system– Immature workload– Leftover workload

Page 99: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Heavy Tails

Page 100: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Tail Types

When a distribution has mean m, what is the distribution of samples that are larger than x?

• Light: expected to be smaller than x+m

• Memoryless: expected to be x+m

• Heavy: expected to be larger than x+m

Page 101: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Formal Definition

Tail decays according to a power law

Test: log-log complementary distribution

20Pr axxXxF a

xaxF log)(log

Page 102: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Consequences

• Large deviations from the mean are realistic

• Mass disparity– small fraction of samples responsible for large

part of total mass– Most samples together account for negligible

part of mass

Crovella, JSSPP 2001

Page 103: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Unix File Sizes Survey, 1993

Page 104: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Unix File Sizes LLCD

Page 105: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Consequences

• Large deviations from the mean are realistic

• Mass disparity– small fraction of samples responsible for large

part of total mass– Most samples together account for negligible

part of mass

• Infinite moments– For mean is undefined– For variance is undefined

1a2a

Crovella, JSSPP 2001

Page 106: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Pareto Distribution

With parameter the density is proportional to

The expectation is then

i.e. it grows with the number of samples

1a2x

xcdxx

cxxE ln1

][2

Page 107: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Pareto Samples

Page 108: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Pareto Samples

Page 109: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Pareto Samples

Page 110: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Effect of Samples from Tail

• In simulation:– A single sample may dominate results– Example: response times of processes

• In analysis:– Average long-term behavior may never happen

in practice

Page 111: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Real Life

• Data samples are necessarily bounded

• The question is how to generalize to the model distribution– Arbitrary truncation– Lognormal or phase-type distributions– Something in between

Page 112: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Solution 1: Truncation

• Postulate an upper bound on the distribution

• Question: where to put the upper bound

• Probably OK for qualitative analysis

• May be problematic for quantitative simulations

Page 113: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Solution 2: Model the Sample

• Approximate the empirical distribution using a mixture of exponentials (e.g. phase-type distributions)

• In particular, exponential decay beyond highest sample

• In some cases, a lognormal distribution provides a good fit

• Good for mathematical analysis

Page 114: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Solution 3: Dynamic

• Place an upper bound on the distribution

• Location of bound depends on total number of samples required

• Example:

Note: does not change during simulation

NFB 211

Page 115: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Self Similarity

Page 116: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

The Phenomenon

• The whole has the same structure as certain parts

• Example: fractals

Page 117: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University
Page 118: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

The Phenomenon

• The whole has the same structure as certain parts

• Example: fractals

• In workloads: burstiness at many different time scales

Note: relates to a time series

Page 119: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Job Arrivals to SDSC Paragon

Page 120: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Process Arrivals to SDSC Paragon

Page 121: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Long-Range Correlation

• A burst of activity implies that values in the time series are correlated

• A burst covering a large time frame implies correlation over a long range

• This is contrary to assumptions about the independence of samples

Page 122: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Aggregation

• Replace each subsequence of m consecutive values by their mean

• If self-similar, the new series will have statistical properties that are similar to the original (i.e. bursty)

• If independent, will tend to average out

Page 123: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Poisson Arrivals

Page 124: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Tests

• Essentially based on the burstiness-retaining nature of aggregation

• Rescaled range (R/s) metric: the range (sum) of n samples as a function of n

Page 125: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

R/s Metric

Page 126: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Tests

• Essentially based on the burstiness-retaining nature of aggregation

• Rescaled range (R/s) metric: the range (sum) of n samples as a function of n

• Variance-time metric: the variance of an aggregated time series as a function of the aggregation level

Page 127: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Variance Time Metric

Page 128: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Modeling Self Similarity

• Generate workload by an on-off process– During on period, generate work at steady pace– During off period to nothing

• On and off period lengths are heavy tailed

• Multiplex many such sources

• Leads to long-range correlation

Page 129: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Research Areas

Page 130: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Effect of Users

• Workload is generated by users

• Human users do not behave like a random sampling process– Feedback based on system performance– Repetitive working patterns

Page 131: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Feedback

• User population is finite• Users back off when performance is

inadequate

Negative feedbackBetter system stability

• Need to explicitly model this behavior

Page 132: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Locality of Sampling

• Users display different levels of activity at different times

• At any given time, only a small subset of users is active

Page 133: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Active Users

Page 134: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Locality of Sampling

• Users display different levels of activity at different times

• At any given time, only a small subset of users is active

• These users repeatedly do the same thing

• Workload observed by system is not a random sample from long-term distribution

Page 135: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

SDSC Paragon Data

Page 136: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

SDSC Paragon Data

Page 137: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Growing Variability

Page 138: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

SDSC Paragon Data

Page 139: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

SDSC Paragon Data

Page 140: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Locality of Sampling

The questions:

• How does this effect the results of performance evaluation?

• Can this be exploited by the system, e.g. by a scheduler?

Page 141: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Hierarchical Workload Models

• Model of user population– Modify load by adding/deleting users

• Model of a single user’s activity– Built-in self similarity using heavy-tailed on/off

times

• Model of application behavior and internal structure– Capture interaction with system attributes

Page 142: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

A Small Problem

• We don’t have data for these models

• Especially for user behavior such as feedback– Need interaction with cognitive scientists

• And for distribution of application types and their parameters– Need detailed instrumentation

Page 143: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Final Words…

Page 144: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

We like to think that we design systems based on solid foundations…

Page 145: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

But beware:

the foundations might be unbased assumptions!

Page 146: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

We should have more “science” in computer science:

• Collect data rather than make assumptions

• Run experiments under different conditions

• Make measurements and observations

• Make predictions and verify them

• Share data and programs to promote good

practices and ensure comparability

Computer Systems are Complex

Page 147: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Advice from the Experts

“Science if built of facts as a house if built of stones. But a collection of facts is no more a science than a heap of stones is a house”

-- Henri Poincaré

Page 148: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Advice from the Experts

“Science if built of facts as a house if built of stones. But a collection of facts is no more a science than a heap of stones is a house”

-- Henri Poincaré

“Everything should be made as simple as possible, but not simpler”

-- Albert Einstein

Page 149: Workload Modeling and its Effect on Performance Evaluation Dror Feitelson Hebrew University

Acknowledgements

• Students: Ahuva Mu’alem, David Talby,

Uri Lublin

• Larry Rudolph / MIT

• Data in Parallel Workloads Archive– Joefon Jann / IBM

– Allen Downey / Welselley

– CTC SP2 log / Steven Hotovy

– SDSC Paragon log / Reagan Moore

– SDSC SP2 log / Victor Hazelwood

– LANL CM-5 log / Curt Canada

– NASA iPSC/860 log / Bill Nitzberg