- We want to develop a sharing method where:
- We want to share only when we don't need the resources ourselves.
- We do not want to share when we do need to use the resources ourselves.
- We want to share only when we don't need the resources ourselves.
- The theoreticians have long ago invented such a sharing
technique.
It is known as an exponential server or the general processor sharing method.
- The General Processor Sharing technique is a purely theoretical
sharing technique where a "processor" (or "server") is shared
in an ideal manner
- Suppose a processor has the capability of processing M
units of work per sec.
- If there is only one job (requiring a certain number of work units
to complete it), then this job is worked on by the processor
at a speed of M units/sec
- On the other hand, if there is two job
(each requiring a certain number of work units to complete it),
then both jobs are worked on simultaneously
by the processor at a speed of M/2 units/sec
- And so on... (so for 3 jobs, all 3 jobs serviced simultaneously
at a speed of M/3 units/sec, etc)
- The exponential server can thus "split" itself into
any number of smaller servers and dedicate one of
the smaller servers to work on a different job.
All jobs will make progress in time, simultaneously.
- The sharing can be generalized by assigning weights
(importance, priority) to jobs - each job will receive
an amount of service that is proportional to its weight.
- Example, if 2 jobs are being worked on and job 1 has weight 1 and job 2 has weight 2, then job 1 will be worked on at a speed of M/3 units/sec while job 2 will be worked on at a speed of 2*M/3 units/sec.
- one "normal" server or processor can work on one job at a time
- hence, only one job will make progress
- Clearly, the exponential server is the canonical (ideal) way
of sharing a processor, a server or a resource
- Depending on the application, approximation techniques can be developed that can approximate the general processor sharing technique.
- This example should be well-known to you and illustrate the fact
that different technique must be developed for different applications
that approximate general processor sharing.
- Given a computer system with 1 CPU
- Given N jobs (programs that need to be executed by the computer)
- How can the jobs share the CPU so that it appears that
all N jobs will run simultaneously
(each at approximately 1/N CPU speed) ?
- The technique used to implemented this is time slicing
and it is implemented using interrupts and
interrupt handling routines.
- I discussed this techinque in CS355 ( click here ), but CS451 should provide more details on this topic.
- Our problem is not sharing a CPU, but a transmission link:
- To make the problem more concrete, let us state it explicitly now....
- Although data transmission is an end-to-end problem,
we will first study the hop-by-hop case:
(End-to-end is the complete path from the sender to the receiver and hop-to-hop is a transfer between any 2 adjacent nodes.)
- Transmission speed of the link is R bps.
- There are N connections or flows that shares a link:
f1,
f2, ...,
fN.
- The reserved data trasnmission rate of the N connections are
r1,bps,
r2 bps, ...,
rN bps.
- How can we make sure that flow f1
receives at least r1 bps,
flow f2
receives at least r2 bps, ... etc
(Note: I say at least because when other flows do not use their reserved portion of the bandwidth (i.e., have no data to send), the bandwidth is given to other flows that do need to it.)
- Let us first handle the unweighted sharing - each flow
receives the same amount of service (data rate)
- This can be done as follows:
- Set up one queue for each flow
- Transmit the data in the queue in a round robin manner
- In each round of round robin, transmit one bit from each flow.
- This particular way of servicing flows is called
the fluid flow service
- A router that implements this type of service is called
a fluid flow server (FFS)
- Note: we can look at "data transmission" as a "service"
(transmission service) and
hence, a router that implements the fluid flow service is called
a fluid flow server
- Note: A weighted version of the FFS can be easily designed. For example, if there are 2 flows and weight of flow 1 is 1 and weight of flow 2 is 4, then in each round, the FFS will transmit 1 bit from flow 1 and 4 bits from flow 4.
- Obviously, the FFS is only an idealized approximation of
the general processor sharing technique
- A real router must transmit a packet atomically
(unsplit) because the structure of the packet must remain
intact.
- The FFS server does not send packets atomically
(mixing bits from a number of packets), and hence it
cannot be used to forward packets.
Nontheless, it gives us an idea what kind a service we aim to design.
In fact, the service discipline that best approximates the General Processor Sharing is designed based on the FFS....