Steven Bakker, DANTE-UK, steven@dante.org.uk
Vincent Berkhart, DANTE-UK, vincent@dante.net
Erik-Jan Bos, SURFnet-Netherlands, bos@surfnet.nl
John Dyer, Terena-EC, dyer@terena.nl
Kilnam Chon, KAIST-Korea, chon@cosmos.kaist.ac.kr
Donal Hanna, Netskills-UK, donal.hanna@ncl.ac.uk
Doug Hughes, Canarie-Canada, dihughes@canarie.ca
Harhisa Ishida, Univ. of Tokyo-Japan, ishida@u-tokyo.ac.jp
Peter Kirstein, Univ. College London-UK, kirstein@cs.ucl.ac.uk
Kazunori Konishi, KDD-Japan, konishi@lab.kdd.co.jp
Steven Ladouceur, Bell Canada, sladouceur@on.bell.ca
Thomas Lenggenhager, Switch-Switzerland, lenggenhager@switch.ch
Mark Luker, National Science Foundation-US, mluker@nsf.gov
John Martin, Terena-EC, martin@terena.nl
Tracie Monk, DynCorp/FNC-US, tmonk@snap.org
Peter Taylor, Mailbase, Newcastle Univ.-UK, peter.taylor@ncl.ac.uk
Tan Tin Wee, National Univ. of Singapore, tinwee@irdu.nus.sg
Duane Wessels, UCSD/NLANR, wessels@nlanr.net
Walter Wiebe, Federal Networking Council-US, wwiebe@nsf.gov

II. ACTION ITEMS

CCIRN members are invited to participate in TERENA’s Caching Task Force meeting in Amsterdam, January 22, 1997. (J. Martin)

Duane Wessels will develop and distribute detailed specifications for configuring caches which will be participating in the global caching hierarchy by January 15th.

Kilnam Chon will provide the CCIRN Caching Working Group participants with information on the Asia Pacific Advanced Networking (APAN) plan for its caching hierarchy and on the results of the Asia Pacific Networking Group (APNG) Caching BoF which is scheduled for January 28, 1997 in Hong Kong.

Participants agreed that there is a great need for new tools to evaluate and predict the performance of caches. Future meetings of this working group will address the specifics of what is required in this area. (CCIRN Co-Chairs)

III. DISCUSSIONS

a. Europe

John Martin of Terena described Terena’s Cooperative Hierarchical Object Indexing and Caching for Europe (CHOICE) project and its current effort entitled Cooperative Hierarchical Object Caching (CHOC) -- described at http://www.terena.nl/tech/projects/choc/.

John explained that the analysis of network traffic on the Internet has shown that significant savings can be made if local caches are introduced and these are linked to form a "caching hierarchy". Much work has been carried out on the development of caching software and many institutions and networks already implement caching at various levels.

The CHOC effort is designed to promote the usage and coordination of caching technologies at the local, national and international level to produce a more efficient, cost-effective and above all reliable service for the users of research networks in Europe. Activities include:

1. Establishment of a TERENA task force for caching deployment and coordination in Europe and coordination with efforts outside Europe. (E.g. DESIRE, NLANR, INSIGHT)
2. Creation and maintenance of a set of documents for those intending to install and operate cache services.
3. Organization of task force meetings to promote and disseminate information about caching and caching coordination.
4. Maintenance of a set of web pages containing information about caching and pointers to caching initiatives in various countries.

Specific activities of the TERENA's Caching task force include:

• Promoting the deployment of caching hierarchies in Europe
• Linking these with initiatives in the US (NLANR)
• Coordinating with other work on caching in Europe (DESIRE, INSIGHT, etc.)
• Providing documentation and services for organizations that want to start caches

CHOC is not intended to:

• configure software/hardware – only to promote cooperation
• determine a country's caching policy
• enforce any policies -- only to advise participants

John described several challenges facing Europe’s caching hierarchy, including:

• Transit traffic for international parenting.
• Tracking bilateral international interconnections (transition towards TEN-34). Historically, Europanet and Ebone served as the backbone providers for most connections. TEN-34 will be operational during the first quarter of 1997. Its use could potentially enable the development of a strict hierarchy via a single route to US. Caching is also being considered as a potential service to be offered over TEN-34.
• The role of the CIX peering point in Northern Europe in Europe's caching hierarchy is unclear.

A survey of caching use in Europe was conducted in the Fall 1996. The detailed analysis of the 200+ responses is available at http://w3cache.icm.edu.pl/survey/results.

The Caching task force mail list is available by sending a message of "subscribe tf-cache [Your Real Name]" to mailserver@terena.nl.

The next meeting of the Caching Task Force (TF-Cache) is scheduled for Amsterdam on January 22, 1996. CCIRN members are welcome to participate.

b. North America

Duane Wessels of UCSD/NLANR described an effort supported by the National Science Foundation to develop and deploy a prototype of a global web caching hierarchy. With root caches located at each node of the NSF's very high speed Backbone Network Service (vBNS), the NLANR web caching hierarchy has experienced a steady increase in usage since its inception in December 1995. Currently, NLANR receives requests from over 100 cache clients in dozens of countries and serves between 15-20 Gbytes per day.

NLANR caches run on DEC Alphaserver 1000 systems at each of the five NSF-sponsored supercomputer center sites (which are connected via the vBNS), and at NASA’s Ames Internet eXchange (AIX) in California.

Duane described the importance of the global caching hierarchy mesh and of research into country-country peering. He noted that pressure to use caches is coming from both the client and server sides, however, thus far, participants of the global caching hierarchy are mostly from the research and education communities. The involvement of commercial providers in Australia and South Africa is the exception to this rule. He also noted that DFN, which has 11 caching nodes in Germany, will soon implement a cache in the U.S. connected to its 155 MB trans-Atlantic connection.

Duane reviewed the NLANR caches log files from 12/11/96. During this 24 hour period NLANR hit rates averaged between 9-20% -- 50% hit rates for these caches would be a goal. As a rule of thumb, bandwidth savings accorded to each cache hit item is estimated at approximately 16Kb.

The goals which the global caching hierarchy might aspire to include:

• having at least one web cache per autonomous system (AS)
• if these caches existed, they might use an AS-based routing database to forward Web requests through caches [the next version of http will permit actual pointing to where a document is to be retrieved from]
• might also have one multicast group per AS – with a port number as an AS number; each listens to popular multicast site

To do this, one would need a database of cider blocks using tools produced by the PRIDE project

Duane reviewed "Excuses and Issues" relating to utilizing caches. These include the following:

• Copyright
  • confusing because real issue is revenue, not necessarily copyrights
  • HTTP provides numerous mechanisms for information provides to prevent or control caching
• Privacy
  • cache requests will most likely be logged somewhere
  • some countries (e.g., Germany) have laws to protect privacy
• Request routing problem related to iTLDs
  • NLANR is currently researching ways to fix these routing problems
• Returning access counts to providers
  • no solutions have yet been implemented
  • there is a draft proposal in the HTTP working group on how to count hits
• Access controls
  • there is a strong desire to protect and control access to data and resources
  • this is at odds with needs for easier configuration

1. developing animated visualizations of how the architecture has changed over time
2. mapping and characterizing paths of the unicast topology underneath each neighbor path
3. performing analysis of if-modified-since requests (what percent of requests are IMS requests, how many of those receive a not modified, should the cache handle IMS requests directly or to pass them to a parent cache?)
4. assessing the tradeoffs of push caching
5. developing mechanisms for server administrators to obtain statistics on hit counts for cache objects from their server
6. developing a better understanding of optimal document expiration values

Information on the NLANR caching project is available at http://www.nlanr.net/Cache. The mail list is available for subscription at ircache-request@nlanr.net.

Doug Hughes and Steven Ladouceur explained that Canada’s Canarie Network (http://www.canarie.ca/ntn) is working to implement a national caching hierarchy for Canada during 1997. Currently there are caches in Montreal, Vancouver, and Toronto. Connections will include three 45 Mbps links to U.S. and an OC3 to Germany by the first quarter of 1997. Canarie will also be interconnected with the vBNS during 1997, allowing direct interface with the NLANR cache hierarchy.

c. Asia

Kilnam Chon explained that many countries in Asia utilize caching, including Australia, Japan, Korea, New Zealand, and Singapore. Indexing and directory services are also emerging. Currently, one Korean provider is reporting hit rates on its caches of upwards 70%. Prof. Goto commented that hit rates in Japan are often around 65%.

A caching hierarchy is currently being employed in Korea. According to Chon, its use saves approximately 350 Mbps total bandwidth or almost $2 million per year. In Japan, there are many cache servers, however, a hierarchy will not be implemented until the Asia Pacific Advanced Networking (APAN) initiative is active in mid 1997.

Under the APAN initiative, Tokyo, Japan and Seoul, Korea will host the root caches for Asia. A 2-4 MB link that would serve as top of hierarchy, connected to the high performance link to the U.S.

d. Discussion

Participants discussed the need for a common policy for default values on features such as refresh rates. This is particularly important for the root caches, but is also needed at higher levels of the hierarchy. There should be different values for different objects types. Participants agreed that a small group of individuals should work to develop these recommended guidelines. Duane Wessels agreed to take the lead, and would provide CCIRN Caching working group members with a draft of these specifications by January 15th.

The next step might include surveying end-users and administrators regarding appropriate settings. Sensitivity analysis is also necessary concerning these settings, e.g. what is the probability of receiving a stale page given a certain default setting. Time-out values and refresh data rates are also important.

There is a great need for new tools to evaluate the performance of a cache. Currently, analysis is done by manually reviewing log files. Predictive tools are also needed.

Preloading information into caches was also discussed. Prefetch (Microsoft), Wcall (Japan), and HPPTGet are among the various packages discussed. Note that pre-fetching popular web pages can increase hits from 40-60%.

Last updated on 23 January 1997