INTRODUCTION

TO

The Workstation Model

Disk Usage	Advantages	Disadvantages
(diskless)	Low cost; easy hardware and software maint.; symmetry and flexibility	Heavy network usage; bottlenecks on file servers
Paging, temporary files	Reduces network load	High cost to purchase disks
Paging, temporary files, binaries	Reduces network load even more	High cost; updating binaries
Paging, temporary files, binaries, caching	Reduces network load even more; reduces load on file server	High cost; cache consistency
Full local file system	V. low network load; no need for file servers	Loss of transparency

Locating an idle processor

Issues

What is an idle workstation?

A workstation is idle if no user-initated processes are running and if input devices have not been touched for several minutes - a machine that is idle can still have a heavy load
How is an idle workstation found?

Idle workstation can either be hunted for (client driven) or can announce its availability (server driven)
How can a remote process be run transparently?

The process may need data from the local environment, local files, etc. But temp. files can be created remotely. What about input from keyboard and output to monitor?
What happens if machine's "owner" comes back?

Do nothing, but what about performance for owner?

Kill the cuckoo, but what about work done?

Migrate the cuckoo - hard, hard, hard!

Always leave the machine as you found it...

The Processor Pool Model

For workstation model, question is when is process executed locally and when remotely - and on which remote machine
For processor pool, question is which processor to use for an incoming process

General strategies

Design Issues for processor alocation algorithms

Example algorithms...

The up-down algorithm

Up-Down tries to distribute CPU time fairly, rather than maximising CPU usage
Coordinator manages centralised usage table containing an entry for each workstation
Whenever a CPU performs a significant scheduling event, a message is sent to the coordinator
Upon creation of a new process, if a CPU doesn't want to run it locally, it asks coordinator to allocate another CPU to the process
How process is scheduled depends on the recent past history of the workstation
For every second of borrowed CPU time, the usage table entry accumulates penalty points
Whenever the coordinator is unable to allocation a CPU to a process, points are deducted from the usage table entry
Processes in the (shared) ready queue are ordered according to the number of points they have accumulated
The process with the least number of points will be serviced first

Sender-Initiated Distributed Heuristic Algorithms

A. Sender picks a machine at random and sends process there. If receiver cannot handle it, it randomly picks another machine, etc., until some machine accepts it or a hop count is exceeded
B. Sender picks a machine at random and sends a probe to enquire about its availability. Continues until machine accepts process or probe limit is exceeded
C. Sender probes k machines for their loads. Sends process to machine with lowest load.

Receiver-Initiated Distributed Heuristic Algorithm

When a machine is idle, it sends a probe to a randomly chosen computer to ask for work
Continues until somebody sends it work, or timer lapses
Better than previous algo when system is overloaded

Not much to say, because scheduling is local
However, sometimes there are problems when two processes (running on different CPUs) are closely related (need to communicate)
Optimal performance is obtained when both processes are run at the same time (otherwise, performance can degenerate to worse than that of uniprocessor systems)