doi:10.1016/j.peva.2007.06.016
Copyright © 2007 Elsevier Ltd All rights reserved.
Performance impacts of autocorrelated flows in multi-tiered systems
aCollege of William and Mary, Williamsburg, VA, USA
bMicrosoft Corporation, Redmond, WA, USA
cSeagate Research, Pittsburgh, PA, USA
Available online 28 June 2007.
Abstract
This paper presents an analysis of the performance effects of burstiness in multi-tiered systems. We introduce a compact characterization of burstiness based on autocorrelation that can be used in capacity planning, performance prediction, and admission control. We show that if autocorrelation exists either in the arrival or the service process of any of the tiers in a multi-tiered system, then autocorrelation propagates to all tiers of the system. We also observe the surprising result that in spite of the fact that the bottleneck resource in the system is far from saturation and that the measured throughput and utilizations of other resources are also modest, user response times are very high. When autocorrelation is not considered, this underutilization of resources falsely indicates that the system can sustain higher capacities.
We examine the behavior of a small queuing system that helps us understand this counter-intuitive behavior and quantify the performance degradation that originates from autocorrelated flows. We present a case study in an experimental multi-tiered Internet server and devise a model to capture the observed behavior. Our evaluation indicates that the model is in excellent agreement with experimental results and captures the propagation of autocorrelation in the multi-tiered system and resulting performance trends. Finally, we analyze an admission control algorithm that takes autocorrelation into account and improves performance by reducing the long tail of the response time distribution.
Keywords: Multi-tiered systems; Autocorrelation; Capacity planning; Workload characterization; Queuing networks
Fig. 1. ACF of inter-arrival and service times for disk level traces measured in enterprise systems and consumer electronics devices.
Fig. 2. A closed system with M queues.
(a) Scenario 1: /Exponential/1 → /MMPP(2)/1.
(b) Scenario 2: /MMPP(2)/1 → /Exponential/1.
Fig. 3. The ACF of departures from Q1 (arrivals to Q2), departures from Q2 (arrivals to Q1) for both scenarios. The ACF of the service process that generates autocorrelated flows in the system is also illustrated.
Fig. 4. Performance measures: (a) mean round trip time, (b) mean queue length, (c) mean utilization, and (d) mean throughput at each queue for Scenario 1. In all experiments the service time in Q1 (non-bottleneck queue) is exponentially distributed. NOACF indicates that Q2 has independent service times. ACF indicates that the Q2 has autocorrelated service times.
Fig. 5. Performance measures: (a) mean round trip time, (b) mean queue length, (c) mean utilization, and (d) mean throughput at each queue for Scenario 2. In all experiments the service time in Q2 (bottleneck queue) is exponentially distributed. NOACF indicates that Q1 has independent service times, ACF indicates that Q1 has autocorrelated service times.
(a) Queue 1 response time, MPL = 25.
(b) Queue 2 response time, MPL = 25.
(c) Round trip time, MPL = 25.
Fig. 6. CDFs of (a) response time at Q1, (b) response time at Q2, and (c) round trip time for Scenario 1 with MPL=25. Q2 remains the bottleneck queue and Q1 is exponentially distributed. The bottleneck queue Q2 has autocorrelated service times in the experiment labeled ACF, and has independent service times in the experiment labeled NOACF.
(a) Queue 1 response time, MPL = 25.
(b) Queue 2 response time, MPL = 25.
(c) Round trip time, MPL = 25.
Fig. 7. CDFs of (a) response time at Q1, (b) response time at Q2, and (c) round trip time for Scenario 2 with MPL=25. Q2 remains the bottleneck queue and its service process is exponentially distributed. The non-bottleneck queue Q1 has autocorrelated service times in the experiment labeled ACF, and has independent service times in the experiment labeled NOACF.
Fig. 8. TPC-W experimental environment.
(a) 128 EBs.
(b) 384 EBs.
(c) 512 EBs.
Fig. 9. ACF at various points in the system. Experiments are done using the browsing mix, a database with 10,000 items, and (a) 128 EBs, (b) 384 EBs, and (c) 512 EBs.
(a) Average response time.
(b) Average queue length.
(c) Average utilization.
Fig. 10. Average performance measures with the browsing mix.
Fig. 11. A queuing model of TPC-W.
(a) MPL = 128.
(b) MPL = 384.
(c) MPL = 512.
Fig. 12. Autocorrelation propagation in our queuing model parameterized using the measurements of Section 4.1 with MPL equal to (a) 128, (b) 384, and (c) 512.
(I) ACF model (successful match).
(II) No ACF model (unsuccessful match).
Fig. 13. Model prediction and experimental performance measures, where the service processes at the front-end server in the proposed model are (I) autocorrelated and (II) uncorrelated.
(a) Round trip time, ACF in front server.
(b) Front server response time, ACF in front server.
(c) DB server response time, ACF in front server.
Fig. 14. CCDFs of (a) round trip time, (b) response time of front server, (c) response time of database server using the model of Section 4.2 where the front server has ACF in its service process. In all experiments MPL is equal to 512.
(a) Round trip time, ACF in DB server.
(b) Front server response time, ACF in DB server.
(c) DB server response time, ACF in DB server.
Fig. 15. CCDFs of (a) round trip time, (b) response time of front server, (c) response time of database server when the database server has ACF in its service process. In all experiments MPL is equal to 512.
(a) Scenario 1.
(b) Scenario 2.
Fig. 16. The mean system utilization at each queue and the system throughput for (a) Scenario 1 and (b) Scenario 2.
Table 1.
Summary of the two scenarios
Table 2.
Hardware components of the TPC-W system
Corresponding address: College of William and Mary, Computer Science, Williamsburg, VA 23185, USA.
Abstract
Purchase PDF (2833 K)
E-mail Article
Add to my Quick Links
Cited By in Scopus (4)
Purchase PDF (515 K)
, Qi Zhangb, 
