Directing a Datacenter: Predicting Resource Utilization from Workload

1
Directing a DatacenterPredicting Resource
Utilization from Workload

• Peter Bodík
• RAD Lab, UC Berkeley

2
Overview slide
3
Introduction

• resource allocation main task of the director
• satisfy SLA reqs, but dont waste
  power/resources
• minimize cost
• need to model/predict resource utilization
• depends on workload, changes w/ app
• modeling resource utilization
• input workload
• output CPU utilization, disk, network, memory
  activity
• this talk 3 models
• linear regression
• quantile regression
• sampling method

4
The datacenter cost function
workload
proposed physical config
input
per-tier resource utilization
server resource utilization
performance, availability
Jeff Chen
server power consumption
server temperature
HVAC power consumption
cost
output
5
Experimental setup

• Rubis PHP application, like eBay
• 2 tiers Apache/PHP and MySQL, running on VMWare
  ESX
• just one web-server
• 27 types of requests, but only using 10
• 3 HTML files, 7 PHP (dynamic) request
• no writes, all reads cached in memory
• measure workload, metrics every 20 seconds
• workload model
• vector of request rates for 10 classes of
  requests
• workload generation
• not using Rubis workload generator
• use exponential interarrival times
• we use interarrivals from World Cup 98 and
  Ebates.com traces

6
Model 1 linear regression

• assumption utilization is a linear function of
  workload
• 1 request -gt 1M CPU cycles, 10 requests -gt 10M
  cycles
• same for disk, network, memory
• reasonable when requests independent
• web server, Ruby on Rails, MySQL
• linear regression
• train and evaluate on same data 2-4 error (CPU
  and net)
• extrapolation two experiments
• increase workload magnitude
• different workload

7
Increasing workload magnitude

• training evaluation
• used 10 different workloads
• web CPU 3 - 10
• larger increase in magnitude -gt larger error
• web network, DB net/CPU 3 - 4

request rates
8
Changing workload

• train on one workload, evaluate on a different
  workload
• training evaluation
• results 6 - 8 error
• web server and DB, net and CPU

9
What else do we need?

• whats the variance of the resource utilization?
• mean not really useful
• whats the error of the predictions?
• can compute, but assuming Normal distribution
• how much data do we need?
• whats the effect of workload variation?

10
Model 2 quantile regression

• estimate 95th percentile, not mean
• formulate as linear program
• assumption P( cpu w ) same for different
  workloads

11
Estimating error -- bootstrap

• resample the data, fit QR, repeat 100x
• estimate error from bootstrap samples

12
Estimating error (2)
13
Just 40 points -- larger error
14
How much data do we need?

• use 95th perc 2stdev as prediction
• model gets more accurate with more data

15
Model 3 sampling

• previous models estimated CPU at a fixed
  workload
• however, workload fluctuates
• need to estimate
• workload distribution
• P( cpu w )
• generate samples of CPU
• sample workload w
• sample cpu P( cpu w )
• repeat
• compute sample 95th percentile

CPU utilization
request rate
16
Estimating workload distribution
17
Estimating P(CPUw)
18
Sampling workload and CPU
19
Results

• actual 95th percentile 347.5 - 4.2 (stdev)
• quantile regression 339.8 - 2.6
• workload and CPU sampling 345.3 - 3.9
• quantile regression already pretty close
• but sampling can accommodate any workload
  fluctuation

20
Comparison of algorithms

• linear regression
• assumptions error from Normal distn
• ignores workload fluctuation
• running time fast
• quantile regression
• assumptions arbitrary error distn, same for
  different workloads
• ignores workload fluctuation
• running time fast, but need bootstrap to
  estimate error
• workload CPU sampling
• assumption arbitrary error distn, same for
  different workloads
• running time slower, sampling bootstrap to
  estimate error

21
Future work

• model lifecycle
• when to retrain the model?
• workload characterization
• real web app 100s - 1000s request classes
• what are the important request classes?
• modeling response time, non-linear resource
  utilization

22
Summary

• use algorithms that make few assumption
• linear regression simple to analyze, but not
  useful in practice
• similar to M/M/1 queue
• quantile regression, sampling
• harder to analyze/estimate
• but work for broader class of web apps and
  workloads
• estimate the prediction error of the model
• model still useful even with large error