Identifying Performance Bottlenecks in CDNs through TCP ...mfreed/docs/wand-wmust11-slides.pdf ·...

Identifying Performance Bottlenecks in CDNs through

TCP-Level Monitoring

Peng Sun Minlan Yu, Michael J. Freedman, Jennifer Rexford

Princeton University

August 19, 2011

Performance Bottlenecks

2

Server

APP

Server OS

CDN Servers

Internet Clients

APP Write too slowly

Server OS Insufficient send buffer or Small initial congestion window

Internet Network congestion

Client Insufficient receive buffer

Reaction to Each Bottleneck

3

Server

APP

Server OS

CDN Servers

Internet Clients

APP is bottleneck: Debug application

Server OS is bottleneck: Tune buffer size, or upgrade server

Internet is bottleneck: Circumvent the congested part of network

Client is bottleneck: Notify client to change

Server

APP

Packet

Sniffer Server OS

Previous Techniques Not Enough

4

Application logs: No details of network activities

Packet sniffing: Expensive to capture

Active probing: Extra load on network

Transport-layer stats: Directly reveal perf. bottlenecks

How TCP Stats Reveal Bottlenecks

CDN Servers Internet Clients

CDN Server Applications

Server Network Stack Network Path Clients

5

Insufficient data in send buffer

Send buffer full or Initial congestion window too small

Packet loss Receive

window too small

Measurement Framework

•  Collect TCP statistics •  Web100 kernel patch •  Extract useful TCP stats for analyzing perf.

•  Analysis tool •  Bottleneck classifier for individual connections

•  Cross-connection correlation at AS level •  Map conn. to AS based on RouteView

•  Correlate bottlenecks to drive CDN decisions

6

How Bottleneck Classifier Works

7

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 180

50

100

150

200

250

300

350

Time in seconds

KB

RwinCwinBytesInSndBuf

BytesInSndBuf = Rwin

Rwin limits sending

Client is bottleneck

Cwin drops greatly and Packet loss

Network path is bottleneck

Small initial Cwin

Slow start limits perf.

Network Stack is bottleneck

CoralCDN Experiment

•  CoralCDN serves 1 million clients per day

•  Experiment Environment •  Deployment: A Clemson PlanetLab node •  Polling interval: 50 ms •  Traces to Show: Feb 19th – 25th 2011 •  Total # of Conn.: 209K •  After removing

Cache-Miss Conn.: 137K (Total 2008 ASes)

•  Log Space overhead •  < 200MB per Coral server per day

8

What are Major Bottleneck for Individual Clients?

•  We calculate the fraction of time that the connection is under each bottleneck in lifetime

9

Bottlenecks % of Conn. With Bottleneck for

>40% of Lifetime

Server Application 10.75%

Server Network Stack 18.72%

Network Path 3.94%

Clients 1.27%

Reasons: Slow CPU or scarce disk resources of the PlanetLab node

Reasons: Congestion window rises too slowly for short conn. (>80% of the connections last <1 second)

Reasons: Spotty network (discussed in next slide) Reasons: Receive buffer too small (Most of them are <30KB) Our suggestion: Use more powerful PlanetLab machines Our suggestion: Use larger initial congestion window Our suggestion: Filter them out of decision making

AS-Level Correlation

•  CDNs make decision at the AS level •  e.g., change server selection for 1.1.1.0/24

•  Explore at the AS level: •  Filter out non-network bottlenecks •  Whether network problems exist

•  Whether the problem is consistent

10

Filtering Out Non-Network Bottlenecks

•  CDNs change server selection if clients have low throughput

•  Non-network factors can limit throughput

•  236 out of 505 low-throughput ASes limited by non-network bottlenecks

•  Filtering is helpful: •  Don’t worry about things CDNs cannot control •  Produce more accurate estimates of perf.

11

Network Problem at AS Level

•  CDN make decision at AS level

•  Whether conn. in the same AS have common network problem

•  For 7.1% of the ASes, half of conn. have >10% packet loss rate

•  Network problems are significant at the AS level

12

Consistent Packet Loss of AS

•  CDNs care about predictive value of measurement

•  Analyze the variance of average packet loss rates •  Each epoch (1 min) has nonzero average loss rate •  Loss rate is consistent across epochs

(standard deviation < mean)

13

Analysis Length # of ASes with

Consistent Packet Loss

One Week 377 / 2008

One Day (Feb 21st) 122 / 739

One Hour (Feb 21st 18:00~19:00)

19 / 121

Conclusion & Future Work •  Use TCP-level stats to detect performance

bottlenecks

•  Identify major bottlenecks for a production CDN

•  Discuss how to improve CDN’s operation with our tool

•  Future Works •  Automatic and real-time analysis combined into

CDN operation •  Detect the problematic AS on the path •  Combine TCP-level stats with application logs to

debug online services 14

Thanks!

Questions?

15