This paper presents several key techniques for

1
(No Transcript)
2
Introduction

• This paper presents several key techniques for
• Designing Web sites that need to handle large
• request volumes and provide high availability.
• It also gives an overview on how many of
• these techniques were deployed at the official
• Web site for the 1998 Olympic Winter
• Games in Nagano, Japan.

3
Topic of Discussion

• Redundant Hardware and Load Balancing
• Web Server Acceleration
• Efficient Management Of Dynamic Data
• 1998 Olympic games site

4
Topic One Redundant Hardware and Load Balancing

• Redundant Hardware Multiple Server running on
  different computers(143 processor for 1998
  Olympic games )
• Load Balancing
• Round-Robin Domain Name server(RR-DNS) approach
  a single domain name is associate with multiple
  IP address. Clients requests specifying the
  domain name are mapped to servers in round robin
  fashion
• problems
• Server-side caching load imbalance even with
  specified TTL
• Client-side caching load imbalance, lower mean
  loads
• Node failures difficult to provide availability

5

• TCP Routing a node of the cluster server as
  router forwarding client request to server nodes
  in the cluster in round-robin order
• advantages
• Using different load -based algorithms
• Detecting Web server node failure
• Achieving good scalability(when combined with
  RR-DNS)

6
Topic One Redundant Hardware and Load Balancing

• Commercially available TCP router
• IBM, Redware, Resonate and Cisco
• Ciscos LocalDirctor the packets returned from
  server go through the router
• IBMs Network Dispatcher(ND)
• running under an embedded OS can rout 10,000
  HTTP request/sec

7
Topic One Redundant Hardware and Load Balancing

• An embedded OS improves router performance by
  optimizing the TCP communications stack and
  eliminating the scheduler and interrupt
  processing overheads of a general-purpose OS
• allowing requests to be routed with an affinity
  toward specific server. This avoid generate
  multiple session key for the requests encrypted
  using Secure Sockets Layer(SSL)

8
Topic Two Web server Accelerators

• Description web sit cache which servers
  frequently requested pages

9
Topic Two Web server Accelerators

• IBMs Web server Accelerators
• Run under an embedded operating system
• Serve higher request pages(5000pages/sec)
• API allow caching of dynamic pages
• To reduce cache miss overhead persistent TCP
  connections need to be kept between the cache and
  the server
• Operate in one of two modes transparent or
  dynamic

10
Topic Two Web server Accelerators

• Performance Cache on a uniprocessor 200-MHZ
  Power PC, and using least recently used(LRU) for
  cache replacement.

11
Topic Three Efficient Dynamic Data Serving

• Dynamic pages are essential at sites that provide
  frequently changing data, CPU overhead associated
  with repeatedly generating them can cause
  performance bottleneck.
• Caching technique can improve the performance,
  but rise with the problem on determining which
  pages to cache and when they become obsolete.

12
Topic Three Efficient Dynamic Data Serving

• Data Update Propagation(DUP) is developed for
  cache management.
• DUP maintains dependencies between cached objects
  and underlying data
• A trigger monitor program can detect changes of
  data, and system can invalidate or update cache
  objects that are obsolete

13
Topic Three Efficient Dynamic Data Serving

• Dependencies are represented by a directed graph-
  object dependence graph(ODG), wherein a vertex
  usually represents an object or underlying data.
  An edge from a vertex v to another vertex u
  indicates that a change to v also affects u.

14
Topic Three Efficient Dynamic Data Serving

• Interfaces for creating dynamic data
• Interface for invoking server programs that
  create dynamic pages has significant effect on
  performance
• Common gateway interface(CGI) creates a new
  process to handle each request which incurs
  considerable overhead
• FastCGI establishes long-running process to web
  requests, but needs some communication overhead
  between web server and process

15
Topic Three Efficient Dynamic Data Serving

• IBMs GWAPI, Netscapes NSAPI, and Microsofts
  ISAPI as well as Apaches modules all run server
  tasks in separate threads. Unfortunately, these
  interface can be tricky to use in practice, with
  issues such as portability, thread safety, and
  memory management.
• More recent approaches, such as IBMs JSP,
  Microsofts ASP, Java Servlets and Apaches
  mod-perl hide those interface and issues of
  thread safety also provide built-in garbage
  collection, thus ease the creation of program,
  maintenance, and portability.

16
Topic Four 1998 Olympic Games Site

• The 1998 Winter games web sites architecture was
  an outgrowth of experience with 1996 Summer games
  web site.
• A key objective for 1998 site was reduce hits by
  giving clients the information on the home page
  for the current day. Redesign of the pages led to
  at least a three-fold decrease of the hit rate.
  The 1998 server log suggests more than 25 of the
  users found the information with a signal hit.

17
Topic Four 1998 Olympic Games Site

• Site Architecture
• Utilized 13 IBM Scalable Power Parallel(SP2)
  system at four complexes scattered around the
  world, containing 143 processors, 78Gbytes of
  memory, and more than 2.4 Terabytes of disk space
  for high performance and availability. 100
  availability was achieved by using replication
  information and redundant hardware.
• Dynamic pages were created via FastCGI
  interface, and cached using the DUP algorithm.
  Achieving cache hit rates of better than 97. The
  1996 web site without employing DUP, many current
  pages were invalidated in the process to ensure
  all stale pages were removed, but the hit rates
  were around 80.
• Prefetching is another key component in
  achieving near 100 hit rates

18
Topic Four 1998 Olympic Games Site

• System Architecture
• web pages were served from four location
  Schaumburg(4 SP2), Illinois Columbus(3 SP2),
  Ohio Bethesda(3 SP2), Maryland and Tokyo (3
  SP2), Japan.
• Each SP2 composed of 10 RISC/6000 UP and 1
  RISC/6000 8-way SMP. Each UP had 512 Mbyte of
  memory and approximately 18 Gbytes of disk space.
  Each SMP had 1Gbyte of memory and approximately 6
  Gbyte of disk space. Numerous machines at each
  location were also dedicate to maintenance,
  support, file serving, networking, routing, and
  various other functions

19
Topic Four 1998 Olympic Games Site

• Data flow from Nagano to the internet and the
  scoring

20
Topic Four 1998 Olympic Games Site

• Data flow from the master database to the
  internet server

21
Topic Four 1998 Olympic Games Site

• Local load balancing and high availability
• IBMs Network Dispatcher were used as load
  balancers(LB)
• LB servers ran the gated routing daemon which
  configured to advertise IP address as routes to
  the routers via dynamic routing protocol. Each LB
  was assigned a different cost based on if it was
  the primary or secondary server for an IP
  address. The secondary server has higher cost.
• The routers redistributed these routes into the
  network. The LB that was the primary source for
  the address assigned to incoming requests at the
  closest complex. Only if the primary LB were
  down, the request would go to the secondary LB.
• Each LB server was connected to a pool of front
  end web servers dispersed among the SP2 at each
  site. Traffic was distributed among web server
  based on advisors information.

22
(No Transcript)
23
Summary

• Since the 1998 Olympic, these technology has been
  deployed at highly accessed sites including the
  Web sites of 1999 Olympic and 1999 Wimbledon
  tennis
• The 1999 Wimbledon site made extensive use of Web
  server acceleration technology that was not ready
  for the 1998 Olympic site
• 1999 Olympic site receive 942 million hits over
  14 days. Peak hit rates of 430,000/min, 125
  million/day. 1998 Olympic site receive 643.7
  million request over 16 days with peak hit rates
  of 110,000/min and 57 million/day.