Sipat Triukose, Zhihua Wen, Michael Rabinovich WWW 2011 Presented by Ye Tian for Course CS05112

Overview Background and Motivation Methodology Performance of Akamai CDN Performance of Consolidated Akamai CDN Review

Background A CDN operator deploys a large number of edge servers throughout the Internet. “outsources” certain hostnames to the CDN’s domain name service

HTTP End-user cnn.com (content provider) DNS root server 12 Nearby Akamai cluster GET index. html 4 HTTP Akamai cluster Akamai global DNS server Akamai regional DNS server

HTTP End-user cnn.com (content provider) DNS root server 12 Nearby Akamai cluster 5 DNS lookup cache.cnn.com Akamai cluster 3 4 ALIAS: g.akamai.net Akamai global DNS server Akamai regional DNS server

HTTP End-user cnn.com (content provider) DNS root server 12 Akamai global DNS server Akamai regional DNS server Nearby Akamai cluster 6 Akamai cluster ALIAS a73.g.akamai.net DNS lookup g.akamai.net

HTTP End-user cnn.com (content provider) DNS root server 12 Akamai global DNS server Akamai regional DNS server Nearby Akamai cluster 7 Akamai cluster DNS a73.g.akamai.net Address

HTTP End-user cnn.com (content provider) DNS root server 12 Akamai global DNS server Akamai regional DNS server Nearby Akamai cluster 8 Akamai cluster GET /foo.jpg Host: cache.cnn.com

HTTP End-user cnn.com (content provider) DNS root server 12 Akamai global DNS server Akamai regional DNS server Nearby Akamai cluster 9 Akamai cluster GET /foo.jpg Host: cache.cnn.com GET foo.jpg

HTTP End-user cnn.com (content provider) DNS root server 12 Akamai global DNS server Akamai regional DNS server Nearby Akamai cluster 10 Akamai cluster

Background Example: outsourcing “images.firm-x.com” Step 1-2: firm-x’s DNS server redirects DNS queries for outsourced hostnames to CDN by returning a CNAME; e.g., images.firm-x.com.CDNname.net Step 3: The user resolves the canonical name, with a query that will arrive at the DNS system for the CDN-name.net domain, operated by the CDN Step 4-5: selects an appropriate edge server (typically close to the client as long as server) and responds to the query with the selected server IP address, enacting the content download from the selected server

Example nslookup “ Example AExample B Sourceplanetlab9.millennium.be rkeley.edu planetlab2.dit.upm.es Source locationBerkeley, CA, USAMadrid, Spain Local DNS CNAMEa1105.b.akamai.net Edge Server IP Edge Server LocationSan Francisco, CA, USAMadrid, Spain

From Berkeley From Madrid

Motivation Two CDN design approaches Akamai-like design: Akamai platform in 2007 spanned more than 3,000 locations in over 750 cities in over 70 countries, with each location having on average less than ten servers Digital Island-like design: utilizes massive data centers, comprising thousands of servers, but in many fewer locations. Limelight currently lists 20 data centers on its web site Question: How many locations is enough from the client- observed performance perspective?

Overview Background and Motivation Methodology Performance of Akamai CDN Performance of Consolidated Akamai CDN Review

Edge Server Discovery Needs to harvest a large collection of edge servers since this requires hostname lookups resolutions from different geographical locations. Utilize DipZoom: Typically more then 400 active MPs, mostly on PlanetLab nodes but also on some academic and residential hosts.

Edge Server Discovery Compiled canonical names outsourced to Akamai of the 95 Akamai customers listed on the Akamai’s Web site. Performed DNS resolution of these names from a large number of network locations over a period of 13 weeks. By clustering these edge servers by city and autonomous system, we conservatively estimate we discovered at least 308 locations.

Youku’s CDN Huajun He, Yang Zhao, Jinfu Wu and Ye Tian, Cost-Aware Capacity Provisioning for Internet Video Streaming CDNs, The Computer Journal (Oxford), 2015.

Overriding CDN Edge Server Selection Why? To measure download performance from a particular edge server rather than the server of Akamai’s choosing. How? Connect to the desired edge server directly using its raw IP address rather than the DNS hostname from the URL.

Controlling Edge Server Caching Why? Compare the download performance from origin server and from the edge server cache. How? Use HTTP’s Cache-Control header. Not working Modern caches use the entire URL strings, including the search string as the cache key. On the other hand, origin servers ignore unexpected search strings in otherwise valid URLs. A request for “foo.jpg?randomstring” will return the valid foo.jpg image from the origin server.

Controlling Edge Server Caching 1, 4, 9 from origin server 2, 3,5, 6, 7, 8 from edge server Precisely control the caching behavior of Akamai’s edge servers by appending search strings to image URLs.

Assessing Client-Side Caching Bias How HTTP download performance from a given Akamai edge server depends on its ping round-trip latency. Perform a set of 4 curl downloads, 5 pings, and a traceroute from each MP to various edge servers. Spearman’s rank correlation coefficient of − indicating a strong dependency.

Overview Background and Motivation Methodology Performance of Akamai CDN Performance of Consolidated Akamai CDN Review

Does a CDN Enhance Performance? Compare download performance Cache vs. non-cache (Note: non-cache is still outsourced) Cache vs. origin CDN delivery outperforms both non-cache and origin delivery in 98% and 96% of cases respectively. Missing penalty can be significant Non-cache: first from CDN edge server to cache; then from cache to the client

How Good Is Akamai Server Selection? How good is the Akamai selection algorithm from the client performance perspective. For each measuring point, the percentage of time that the Akamai-selected server performed better than the alternative server CDNs rarely select the best edge server but successfully avoid the worst ones. In roughly 75% of the MPs, the Akamai-selected server outperformed half of the alternatives.

How Good Is Akamai Server Selection? The average difference, the maximum gain, and the maximum loss of throughput of the Akamai-selected server vs. the alternatives. Maximum gain and maximum loss are the throughput differences of the Akamai- selected server over, resp., the slowest and fastest alternative server. There is a substantial room for further improvements.

Overview Background and Motivation Methodology Performance of Akamai CDN Performance of Consolidated Akamai CDN Review

Data Center Consolidation Group edge servers that are close to each other into hypothetical consolidated data centers Refer to as big clusters. “Place” each big cluster into the location of a central server in the cluster called the representative of the cluster. For a given client, we replace the server selected by Akamai, S akam with the representative of the cluster to which S akam belongs.

Data Center Consolidation Estimate the pairwise network distances between all the servers using the dynamic triangles method Use the so-called hierarchical clustering with complete linkage method, following by the cut-the-tree procedure

DipZoom Experiment From each measurement point, downloaded the object of a given size from the Akamai-selected server and from the center of its cluster in the consolidated platform. Group 10,231 edge servers into 150, 100, 60, 40, and 20 data centers

DipZoom Experiment 150, 100, and 60 data centers: consolidated and current platforms show performance advantage over each other with equal probability. 40 and 20 data centers: the original platform starts outperforming the consolidated configuration

DipZoom Experiment Group the MPs by the maximum download throughput

DipZoom Experiment The existing Akamai configuration with large number of data centers favors well-connected users. Compare the performance of existing vs. consolidated configurations for all measurement points with bandwidth below 1.5Mbps A typical download bandwidth for DSL users.

A Live Study Built a Web page with an AJAX application that measures latency to the server that Akamai selects for this browser as well as to a list of other Akamai edge servers, and reports the results to us. Pick a CNAME (a1694.g.akamai.net) which is mapped by Akamai to 979 edge servers. Partitioned the 979 servers into 10, 50, and 100 clusters.

A Live Study Consider the differences between the measured latencies from each client to its closest consolidated data center and the Akamai-selected server. Negative indicates that the consolidated platform is better Positive reflects existing Akamai platform is better Even with significant consolidation, better server selection can make up for any performance losses from consolidation.

Summary Two CDN design approaches? Example. How CDN redirects its clients to one of its edge server? From user performance perspective, can we consolidate the Akamai CDN platform? To what extent?