Internet: Quality of Service Mechanisms at Application Level

1
Internet Quality of Service Mechanisms at
Application Level

• Special Topics in Communications Networks QOS
• Hanoch Levy
• Computer Science
• Tel-Aviv University
• 1999

2
References

• E. Cohen B. Krishnamurthy, J. Rexford, Efficient
  Algorithms for Predicting Requests to Web
  Servers, http//www.research.att.com/edith
• E. Cohen B. Krishnamurthy, J. Rexford,
  Evaluating Server-Assisted Cache Replacement in
  the Web, http//www.research.att.com/edith
• E. Cohen, H. Kaplan, Reducing User-Perceived
  Latency by Prefetching Connections and
  Pre-warning Servers, http//www.research.att.com/
  edith
• www.netcognition.com
• http//www.cisco.com/warp/public/cc/cisco/mkt/scal
  e/distr/prodlit/dd_ds.htm
• http//cs-www.bu.edu/faculty/crovella others
  in cs-www.bu

3
Internet and Web

• Data traffic (internet mostly) -- grows at
  exponential rates.
• Expected to dominate market (take considerable
  fraction of all traffic).
• Web traffic -- dominant fraction of internet
  traffic.
• Elementary Web operation get page (httpget).
• Translates into a TCP connection and a stream of
  IP packets.

4
Application/Transport QOS

• Control at network level controls located at
  network switches (routers) and discriminate the
  packets to achieve performance objectives.
• Control at application level Controls located at
  special network points and discriminate high
  level objects (web documents) to achieve
  performance objectives.

5
Web transfer chain

• User computer

ISP (proxy)
network
server
User computer
Load balancing
servers
6
Load Balancing (sharing)

• Used for large volume web sites.
• Site data is served by multiple mirrored servers
  (could be multiple locations).
• Load sharing device - Aware of relative load of
  the various servers- Directs new request to the
  lowly loaded server.- Desired keep certain user
  on certain server (due to session continuity and
  management).
• May operate on geographically distributed
  environment (LA NY)
• Make geography based decisions
• ExampleCisco DistributedDirector
  http//www.cisco.com/warp/public/cc/cisco/mkt/scal
  e/distr/prodlit/dd_ds.htm
• Q what is session? How is controlled?

7
Load balancing (cont.)

• ExampleCisco DistributedDirector
  http//www.cisco.com/warp/public/cc/cisco/mkt/scal
  e/distr/prodlit/dd_ds.htm
• Example Radware
• Q what is session? How is controlled?

8
Network Control (load balance)

• Cannot operate at application level.
• Also network efficiency and routing supposedly
  handled at lower layers.
• Yet one can try optimize at application level
• Radware http//www.radware.com/product/lproof/Def
  ault.htm
• Multihoming environment enterprise connection to
  internet is via multiple links (multiple ISPs)
• Done typically for redundancy

9
Network Control (load balance)

• internet

Isp2 (zahav)
Isp1 (netvision)

• TAU

10
Network Control (load balance)

• Problems faced by TAU
• Balance the load between 2 isps (as to reduce
  delay)
• Optimize put requests on their specific faster
  link)

11
ISP (Proxy)

• Proxy sits at ISP premises.
• Receives requests from clients.
• Sends request to Net and receives replies.
• Sends replies to clients.
• Intervene in between.

12
Proxy Intervention

• Security (not in our scope)
• Prioritization and discrimination (Comais talk)
  between users.
• Caching - A large disk is used to store web
  pages that requested in the past.

13
Caching and Prefetching

• Caching keep a copy of a document you think you
  might need in future
• Prefetching Bring and keep a copy of page you
  think you might need in future
• Refreshing For a page residing in cache,
  browser/proxy must check whether copy is still
  fresh enough or stale.

14
Examples

• Dynacache
• CacheFlow
• Inktome
• Others
• Squid - Most popular. Public domain.

15
Sky Caching

• Serves a network of ISP proxies.
• Consists of a central server, connected via
  satellite (and network) to N ISP proxies.
• Proxies report central on elements requested.
• Central broadcasts the freshly popular
  documents to other proxies.
• Example Starr report retrieved from Net by first
  3 proxies. All rest already got it due to
  skycache.

16
Server Side Prefetching

• Basic idea Server knows the best site next
  document statistics.
• Use Java code installed on site web pages (in
  addition to the other text/graphics).
• The code starts running at user client.
• This code retrieves the proper pages from the
  site and stores them at clients local cache.
• Ref www.netcognition.com (Raanana)

17
Example

• Server knows that following page A users are
  likely to request page B.
• On page A server attaches get_url(B) Java code.
• Once A is retrieved, and while user reads it,
  Java code retrieves B (seamlessly).
• When user retrieves B -- page comes from cache.

18
Server Side Prefetching (cont.)

• Advantage server has good knowledge of public
  statistics.
• Disadvantage - In many sites fan-out is
  very large requires site to transmit large
  volume (3x) to gain significant reduction in
  expected delay.

19
Client (Browser)

• Caching - done at browsers (Netscape,
  Explorer). - User can control parameters
  (e.g. cache size, and freshness check
  frequency).
• Prefetching - Add-on packages that reside on PC.
  - At idle times program runs in background
  and brings user popular pages. - Good mainly
  for sites often visited by user - Good mainly
  for relative static sites.

20
Client (cont.)

• Can utilize information on user patterns (e.g.,
  time of coming to office).
• Examples Netaccelerator (search in TUCOWS).

21
Connection Prefetching

• Significant fraction of perceived delay is due to
  connection setup (1-2 secs).
• Includes - Domain name lookup - Opening TCP
  connection
• Idea - Once retrieved page A, browser (or
  proxy) identifies links to external sites. -
  Browser initiates domain name lookup and TCP
  connection to those sites. - Upon user requests
  to those sites -- latency reduction

22
Distributed Cache
23
WAN of caches (Akamai)