Internet Content Delivery over Satellite

1
Internet Content Delivery over Satellite

• NANOG October 1999
• Evan Baer
• SkyCache, Inc.

2
SkyCache Overview

• History The Internet is point to point
• Most data carried on the Internet is broadcast
• Web Content
• Usenet News
• Streaming A/V, FTP data, etc.
• One to many, not one to one
• Satellite technology good fit for broadcast data
• Data is transmitted to entire geographic area
• Scales to an unlimited number of receivers
• SkyCache Transmit Internet data over satellite,
  broadcast, to the edge of the network
• Work in conjunction with existing Internet
  infrastructure
• Increased efficiency of bandwidth
• Move broadcast data to a satellite broadcast
  overlay network

3
SkyCache Architecture

• Distribute Internet data via satellite to the
  edge of the network
• Determine popular web objects by examining cache
  user requests, transmit popular objects to cache
  server at the ISP (reactive caching)
• Deliver Usenet newsfeed to newsserver at the ISP
• Transmit streaming a/v, ftp data, and other
  future data to servers at the ISP
• Internet data collected and transmitted at the
  Master Control (MC)
• MC collects web objects, newsfeed, other data for
  transmission
• Cache miss stream user requests are examined at
  the MC via a rules system to determine web
  objects most likely to be cache hits to the
  SkyCache cache community
• Data is aggregated into MC protocol (MCP) and
  sent to the uplink
• Satellite feed received at the SkyCache Adaptor
  (CA) at the customer site
• CA acts as a protocol engine, receiving MCP from
  the satellite, feeding data via http, nntp, and
  other protocols to servers at the ISP
• Not directly involved in client redirection
• SkyCache is a satellite overlay network
• Client redirection handled by ISP in current
  service offerings
• Services for content providers will require being
  involved in end user redirection

4
Full System Architecture
Customer Point of Presence (POP)
Satellite Receiver
SkyCache Adapter
SkyCache Uplink
GE-3 US GE-1E Europe KU Band
45 Mbs
Fast Ethernet
4 Mbs
Satellite Dish (on roof) 75-ohm RG6 coaxial
cable (approximately. 1Ghz RF signal)
Cache Server
News Server
FTP Server
Streaming A/V
Future Services
Satellite Transmitter
Internet
L4 Switch
Router
Master Control
Access Server
Router
Dedicated / Dial-Up Customers
Remote Server
5
Daily Transmit Rates

• Transmit rates from September 15-22
• Currently 4MB/s feed
• 45 MB/s pipe in testing
• Cache turbocharging feed 6-7GB daily
• Usenet newsfeed 20-25GB daily

6
Reactive Caching

• Internet users typically view the same web
  content, regardless of location
• Mostly true, with some differences for time zone
  and geographic location
• More complex outside of North America (more
  localized in Europe, for instance)
• If the same url is seen from multiple caches, it
  is probably of interest to the entire community
• For any given popular site on the net, someone
  clicks first
• If the url coincidence occurs in a reasonable
  length of time, minutes or hours, not days
• SkyCache rules system detects url coincidences at
  the MC
• Potential win on both ends of web popularity
  spectrum
• Top 10 of sites, which are popular day to day,
  popular pages from these sites are sent as they
  expire from the caches and get re-requested
• Other 90 of sites, of which different sites may
  be popular from one day to the next, popular
  pages from these sites are sent as they are
  detected as popular
• Cacheable content from Yahoo, Wired Magazine,
  etc. which is sent multiple times per day
• Cacheable content from short-term popular sites,
  such as irs.ustreas.gov during the middle of April

7
Preloading Web Caches

• SkyCache CA peers as an ICP sibling with a cache
  to inject objects and listen to cache miss
  requests
• CA sends the cache an ICP query for a given
  object
• Cache then sends the CA an ICP hit or miss, if it
  has the object or not
• If the CA received an ICP miss, the injection
  proceeds, and the CA sends the cache an HTTP GET
  request for the object
• The cache then sends an ICP query to the CA for
  the object
• The CA responds with an ICP hit
• The cache sends the CA an HTTP GET for the
  object, and now has the object in cache

SkyCache CA
Cache
ICP_OP_QUERY
ICP_OP_MISS
or
ICP_OP_HIT
HTTP GET
ICP_OP_QUERY
ICP_OP_HIT
HTTP GET
8
The Cache Miss Stream

• SkyCache CA is queried as an ICP sibling for all
  cache miss requests
• On a cache miss, the cache will send an ICP query
  to the CA for the requested object
• The CA responds with an ICP miss
• The cache retrieves the request from the origin
  server
• The CA forwards these url requests to the MC,
  where they are analyzed
• On a cache hit, the request is served from the
  cache, and the CA is never queried

Origin Web Server
HTTP GET
SkyCache CA
Cache
Cache User
ICP_OP_QUERY
HTTP GET
ICP_OP_MISS
9
Finding popular web objects

• US miss stream from September 15-22
• 15 million urls daily, from 75 unique US cache
  locations
• Cache miss stream is comprised of all cache miss
  requests from all active SkyCache enabled caches
• Popular objects defined as an object which has
  been seen multiple times from multiple caches
• 10-15 of all urls in the miss stream are found
  to be popular
• Nearly equal number of cacheable and noncacheable
  objects

10
Cache Performance Increase

• SkyCache injected objects only make an impact on
  the first hit
• Example 1 cache miss and 10 cache hits on
  www.foo.com, can only convert the first miss to a
  hit
• Unique hits are tallied to determine the most
  benefit which could be made to a caches hit
  rates
• Best suited to small and medium caches
• Cache object hit rate can get to 50-55 with
  SkyCache, past that is more a function of user
  base
• 10,000 users request 25,000 different urls
• 100,000 users request 35,000 different urls, more
  hits on the same objects
• Regional caches, servicing 3M cache requests
  per day, have a cache community to large to make
  use of SkyCaches virtual cache community

11
Examining Unique Hits

• Unique hits for 3 SkyCache enabled caches from
  September 15-22
• 40-50 of unique hits are on SkyCache injected
  objects

12
Increased Cache Object Hit Rates

• Increase in object hit rates on SkyCache enabled
  caches from September 15-22
• 15-20 increase of object hit rate due to hits on
  SkyCache objects
• 8-12 points on top of standard cache hit rate

13
Increased Cache Byte Hit Rates

• Increase in byte hit rate on SkyCache enabled
  caches from September 15-22
• 25-30 increase of byte hit rate due to hits on
  SkyCache objects
• 8-10 points on top of standard cache byte hit
  rate

14
Usenet News

• Usenet Newsfeed delivered over satellite,
  sustained 2.5-3MB/s
• SkyCache CA appears as a NNTP feeder
• Snapshot of 3 customers on September 24th, 1999
• 5 minute throughput rates in megabytes
• Feed rates depend upon other customer news peers,
  many other factors
• Fairly obvious when expirations / other server
  maintenance occurs

15
Issues in satellite delivery to edge servers

• Determing effectiveness of delivered data no
  simple task
• Currently requires post analysis of cache access
  logs
• Cache misses are seen in the current ICP cache
  peering mechanism, not hits
• Newsfeed impact is more straightforward, as news
  was built for feeds
• Pushing protocols to do things they werent
  really intended to do
• NNTP fairly simple, protocol meant to feed
  objects
• Slight abuse of ICP and HTTP to inject objects
  into caches, currently no other standard
  protocols
• Streaming A/V protocols also a bit tricky to
  deliver in broadcast fashion
• For current ISP targeted services, client
  redirection handled by ISP
• Future content provider driven services require
  managing client redirection

16
Future

• Broadcast data not on the Internet as we know it
• Overlay networks to distribute content aimed at
  the masses
• Alternative content delivery becomes more
  interesting and accepted as end users get high
  speed connectivity and become less tolerant of
  high latency
• Content providers will be pushed to distribute
  content by the end user
• Overlay networks, server distribution, and
  caching will all play a role
• Growth of caching as accepted practice, more
  functionality
• Better and more widely supported cache protocols
• Passing of cache usage information to peers and
  content providers
• More information about cacheable sites, cache
  breaking sites on the Net
• Cacheable / Uncacheable web sites continue to be
  balanced
• Some sites will continue to have uncacheable
  pages, some will make their pages cache friendly.
  Not all sites will become dynamic or
  uncacheable, nor will all sites become cacheable.