1. Chapter 8-2 : Multicomputers Multiprocessors vs multicomputers Multiprocessors vs multicomputers Interconnection topologies Interconnection topologies Switching schemes Switching schemes Communication with messages Communication with messages Remote procedure calls Remote procedure calls Distributed shared memory Distributed shared memory Multicomputer scheduling Multicomputer scheduling Virtualization Virtualization 1

2. Multicomputers Multiprocessors share the common memory Multiprocessors share the common memory Multicomputers are tightly coupled CPUs with memories of their own and do not share them Multicomputers are tightly coupled CPUs with memories of their own and do not share them These systems are also known as cluster computers and COWS (Clusters of Workstations) These systems are also known as cluster computers and COWS (Clusters of Workstations) The secret of high performance is the interconnection network The secret of high performance is the interconnection network 2

3. 3 Multiprocessor vs Multicomputer Figure 8-29. Comparison of three kinds of multiple CPU systems. Figure 8-29. Comparison of three kinds of multiple CPU systems.

4. 4 Interconnection Topologies Figure 8-16. Various interconnect topologies. (a) A single switch (b) A ring (c) A grid or mesh (d) A double torus (e) A cube (f) A 4D hypercube Note : A diameter is defined as the longest path between any two nodes Note : A diameter is defined as the longest path between any two nodes

5. Switching Schemes Store-and-forward packet switching : messages are broken into packets. Packets move from one switch to another Store-and-forward packet switching : messages are broken into packets. Packets move from one switch to another This scheme has increasing latency (delay) problem because of store-and-forwarding in intermediate switches This scheme has increasing latency (delay) problem because of store-and-forwarding in intermediate switches Circuit switching : a path is established from the source to the destination. Once this path is setup bits are pumped from source to destination with no buffering in the intermediate switches Circuit switching : a path is established from the source to the destination. Once this path is setup bits are pumped from source to destination with no buffering in the intermediate switches Circuit setup may need some time but then the transfer is fast Circuit setup may need some time but then the transfer is fast 5

6. User Level Communication Software Multicomputers communicate through messages Multicomputers communicate through messages Send(dest, &mptr) : send a message pointed by mptr to a process identified by dest Send(dest, &mptr) : send a message pointed by mptr to a process identified by dest Receive(addr, &mptr) : addr is usually CPU number and a process or port number Receive(addr, &mptr) : addr is usually CPU number and a process or port number Send calls can be blocked or nonblocked Send calls can be blocked or nonblocked Receive calls are always blocked. That is, the receiving process has to wait if the message has not been sent before Receive calls are always blocked. That is, the receiving process has to wait if the message has not been sent before 6

7. 7 Blocking Calls Figure 8-19. (a) A blocking send call.

8. 8 Nonblocking Calls Figure 8-19. (b) A nonblocking send call.

9. 9 Remote Procedure Call Figure 8-20. Steps in making a remote procedure call. The stubs are local procedures implementing the interface between the client and the server The stubs are local procedures implementing the interface between the client and the server Parameter and result passing are done by messages Parameter and result passing are done by messages

10. 10 Distributed Shared Memory (1) Figure 8-21. Various layers where shared memory can be implemented (a) The hardware (b) The operating system (c) User-level software

11. 11 Distributed Shared Memory (2) Figure 8-22. (a) Pages of the address space distributed among four machines. When a CPU references an address that is not local, a trap occurs, and the DSM software fetches the page containing the address and restarts the faulting instruction When a CPU references an address that is not local, a trap occurs, and the DSM software fetches the page containing the address and restarts the faulting instruction

12. 12 Distributed Shared Memory (3) Figure 8-22. (b) Situation after CPU 0 references page 10 and the page is moved there.

13. 13 Distributed Shared Memory (4) Figure 8-22. (c) Situation if page 10 is read only and replication is used.

14. Multicomputer Scheduling On a multiprocessor, all processes reside on the shared memory On a multiprocessor, all processes reside on the shared memory On a multicomputer each node has its own memory and its own set of processes On a multicomputer each node has its own memory and its own set of processes Scheduling within a node is easier but allocation of processes to nodes is much more important (load balancing) Scheduling within a node is easier but allocation of processes to nodes is much more important (load balancing) The algorithms and heuristics for doing this assignment are known as processor allocation algorithms The algorithms and heuristics for doing this assignment are known as processor allocation algorithms 14

15. Scheduling Algorithms Graph-Theoretic Deterministic Algorithm Graph-Theoretic Deterministic Algorithm Sender-Initiated Distributed Heuristic Algorithm Sender-Initiated Distributed Heuristic Algorithm Receiver-Initiated Distributed Heuristic Algorithm Receiver-Initiated Distributed Heuristic Algorithm Please read from the book Please read from the book 15

16. Virtualization Virtualization is the virtual machine concept introduced in Chapter 1 – Introduction Virtualization is the virtual machine concept introduced in Chapter 1 – Introduction VM technology allows a single computer to host multiple virtual machines, each potentially running a different operating system VM technology allows a single computer to host multiple virtual machines, each potentially running a different operating system Advantages : Advantages : Failure in one VM does not automatically bring down any others since VMs are isolated Failure in one VM does not automatically bring down any others since VMs are isolated Fewer machines to provide an environment for different applications (instead of having one server for each of the different apllications, a server with VMs can do the job) Fewer machines to provide an environment for different applications (instead of having one server for each of the different apllications, a server with VMs can do the job) Ability to run different and probably older versions of Oss needed by some appications Ability to run different and probably older versions of Oss needed by some appications 16

17. Virtualization Methods Type 1 hypervisor (virtual machine monitor) Type 1 hypervisor (virtual machine monitor) Type 2 hypervisor Type 2 hypervisor 17