1. Scheduling in Distributed Systems
By Arlington Wilson

2. Presentation content Problem Definition
What is Distributed system scheduling?
What conditions does it apply under?
Ousterhout 3 Co-Scheduling algorithms
Matrix Algorithm
Continuous Algorithm
Undivided Algorithm
Co-scheduling effectiveness analysis
Implicit co-scheduling
References

3. Problem definition
Normally each processor does it own local scheduling.
Processes A & B run on processor 0 and processes C & D run on processor 1.
Processors
Time
slot 1 1 2 3 4 5 A C B D

4. Problem definition
How to improve performance when a group of heavily interactive processes run on different processors?
In 1982, John Ousterhout proposed 3 scheduling techniques based on a concept he called co-scheduling.

5. Co-Scheduling Assumptions
Related processes are started together.
In general intra-group communication is more likely than inter-group communication.
Sufficiently large numbers of processors are available to handle the largest group.
Each processor is multi-programmed with N process slots (N-way multiprogramming)

6. Algorithm 1 - The Matrix Method
Processors
P = # of processors, Q = # of processes/group
Allocation
If Pavail/row 0 >= Q, assign to row 0. If not check row 1, etc...
Scheduling
A round robin mechanism
Time
Slots 1 2 3 4 5 X

7. Algorithm 1 - The Matrix Method
Processors
Alternate selection
If an executing row has a empty or blocked process slot, scan for next slot in column to execute.
Time
Slots 1 2 3 4 5 X

8. Algorithm 1 – The Matrix Method
Advantages:
Simple allocation and scheduling algorithms
Disadvantages:
A group cannot be spread over multiple
rows, thus many processor slots are left
Empty.
Alternate selection causes opportunities for co-
scheduling to me missed.

9. The continuous Algorithm
Allocation
Start a window of size P at the left end of the process sequence
If # of empty slot >= Q, allocate the group.
Else, slide window to the right until left most process slot is empty.

10. The continuous Algorithm
Scheduling
Start a window of size P at the left end of the process sequence
At the beginning of each time slice, until the start of a new process group.
When the window has empty process slots or processes that aren’t running, use the alternate selection mechanism.

11. The Continuous Algorithm
Advantages
Better utilization of processor slots
Disadvantages
When the process sequence becomes populated with non-continuous empty slots, during allocation groups become fragmented.
Fragmentation leads to poor performance.

12. The Undivided Algorithm
Same as the continuous algorithm, except that during allocation all new groups are required to be contiguous in the linear process sequence. (No dividing between slots)
Less efficient at packing than continuous.
Substantial better system behavior under heavy loads due to reduced fragmentation.

13. Analysis of the Algorithms
Results presented in terms of co-scheduling effectiveness.
The ability to co-schedule grouped processes not the performance.
Co-scheduling effectiveness is the mean of the ratio of the total # of processors executing co-scheduled processes to the total # of processors with runnable processes. (Ideal is and effectiveness of 1)

14. What lowers co-scheduling effectiveness?
Straddling – The continuous and undivided algorithm, would straddle the right end of the scheduling window. Thus causing the processes inside the window to execute as a fragment, lowering co-scheduling effectiveness.
Alternate Selection – Creates fragments and will degrade the system co-scheduling effectiveness.

15. Analysis of the Algorithms
For average system loads (<1.0) All three algorithms perform about the same due to low straddling and alternate selection.
For system loads above 1.0 both straddling and alternate selection occur cause effectiveness to degrade.
For heavy loads, the undivided algorithm performs 5-10% better then the continuous algorithm which performs 5-10 percent better than the matrix algorithm.

16. Implicit Co-Scheduling [1]
If a process sends a message, then waits a period of time, if no response, assume other group processes are not executing and relinquish the processor.
If a process receives a message, it can infer that other processes in the application are executing and this process should be scheduled.
It uses no global coordinator.

17. References
Scheduling with implicit Information in Distributed Systems - Andrea C. Arpaci-Dusseau, David E. Culler, Alan Mainwaring – Sigmetric ’98
Scheduling Techniques for Concurrent Sytems – John K. Ousterhout ’80
Distributed Operating Systems – Andrews Tanenbaum