Simulation of computer system serving WWW

A computer system serving WWW requests (a web server) can be modelled at a high level as a network of queues: client requests arrive over the network, are processed by the server's CPU, memory and disk, and responses are sent back through the network interface. The model estimates response time, throughput and utilisation of resources, and helps to size the server.

Components

• Request source: users (clients) generating HTTP requests, with an arrival process (Poisson) or a fixed population with think time.
• Network / connection queue: the listen queue, where connections wait for a server thread.
• CPU: parses requests, runs scripts.
• Memory / cache: a file or page cache; hit means no disk access.
• Disk: serves cache misses.
• Network interface: sends the response (service time depends on file size and bandwidth).

High-level sketch

 Clients --> [Listen queue] --> ( CPU )
 (requests)                        |
                         cache hit?|
                         +--yes----+--no--> [Disk queue]->(Disk)
                         |                         |
                         v                         v
                      [NIC queue] --> (Network out) --> Clients

Model parameters

• Arrival rate (requests/s) and its distribution.
• CPU service time per request, probability of cache hit (for example 0.8), disk service time, file size distribution (heavy tailed), bandwidth.
• Limits: maximum number of threads or connections; requests beyond this are rejected.

Simulation procedure

1. Generate request arrivals with random inter-arrival times.
2. Each request seizes a thread, then queues for the CPU.
3. With probability 0.8 it is a cache hit; otherwise it goes to the disk.
4. It then queues for the network interface and sends the response; the response time is recorded at departure.
5. Repeat for many requests, with warm-up removed.

Outputs

Average and 95th-percentile response time, throughput, utilisation of CPU, disk and network link, queue lengths and number of rejected requests. The analyst can change the model (a faster CPU, a larger cache, more bandwidth) and find the bottleneck resource.