1. TECHNIQUES FOR REDUCING CONSISTENCY- RELATED COMMUNICATION IN DISTRIBUTED SHARED-MEMORY SYSTEMS J. B. Carter University of Utah J. K. Bennett and W. Zwaenepoel Rice University

2. INTRODUCTION Distributed shared memory is a software abstraction allowing a set of workstations connected by a LAN to share a single paged virtual address space Key issue in building a software DSM is minimizing the amount of data communication among the workstation memories

3. Why bother with DSM? Key idea is to build fast parallel computers that –are cheaper than conventional architectures –are convenient to use Conventional parallel computer architecture was the shared memory multiprocessor

4. CPU Shared memory Conventional parallel architecture CACHE CPU

5. Today’s architecture Clusters of workstations are much more cost effective –No need to develop complex bus and cache structures –Can use off-the-shelf networking hardware Gigabit Ethernet Myrinet (1.5 Gb/s) –Can quickly integrate newest microprocessors

6. Limitations of cluster approach Communication within a cluster of workstation is through message passing –Much harder to program than concurrent access to a shared memory Many big programs were written for shared memory architectures –Converting them to a message passing architecture is a nightmare

7. Distributed shared memory DSM = one shared global address space main memories

8. Distributed shared memory DSM makes a cluster of workstations look like a shared memory parallel computer –Easier to write new programs –Easier to port existing programs Key problem is that DSM only provides the illusion of having a shared memory architecture –Data must still move back and forth among the workstations

9. Characterizing a DSM (I) Four important issues: 1. Size of transfer units (level of granularity) Big units are more efficient – Virtual memory pages Can have false sharing whenever page contains different variables that are accessed at the same time by different processors

10. False Sharing accesses x accesses y x y page containing x and y will move back and forth between main memories of workstations

11. Characterizing a DSM (II) 2.Consistency model Strict consistency is not possible Various authors have proposed weak consistency models –Cheaper to implement –Harder to use in a correct fashion

12. Characterizing a DSM (III) 3.Portability of programs Some DSMs allow programs written for a multiprocessor architecture to run on a cluster of workstations without any modifications ( dusty decks ) More efficient DSMs require more changes 4.Portability of DSM Some DSMs require specific OS features

13. MUNIN Developed at Rice University Based on software objects (variables) Uses the processor virtual memory to detect access to the shared objects Includes several techniques for reducing consistency-related communication Only runs on top of V kernel

14. Key features Software release consistency: only requires the memory to be consistent at specific synchronization points, Multiple consistency protocols: allow the user to select the best consistency protocols for each data item, Write-shared protocols: reduce false sharing, An update-with-timeout mechanism

15. SW RELEASE CONSISTENCY (I) Well-written parallel programs use locks to achieve mutual exclusion when they access shared variables –P(&mutex) and V(&mutex) –lock(&csect) and unlock(&csect) –request ( ) and release( ) Unprotected accesses can produce unpredictable results

16. SW RELEASE CONSISTENCY (II) SW release consistency will only guarantee correctness of operations within a request/release pair No need to propagate new values of shared variables until the release Must guarantee that workstation has received the most recent values of all shared variables when it completes a request

17. SW RELEASE CONSISTENCY (III) shared int x; request( ); x = 1; release ( ); // propagate x= 1 shared int x; request( ); // wait for new value of x x++; release ( ); // propagate x= 2

18. SW RELEASE CONSISTENCY (IV) Munin uses eager release : new values of shared variables are propagated at release time – Lazy release delays propagation until a request is issued ( Threadmarks ) A workstation issuing a request gets the current values of all shared variables –Shared variables are not associated to a particular critical section (as in Midway )

19. Munin Implementation (I) Three kinds of variables: 1.Ordinary variables: can only be accessed by the process that created them 2.Shared data variables: should always be accessed from within critical regions 3.Synchronization variables : locks, barriers or condition variables must be accessed through special library procedures.

20. Munin Implementation (II) When a processor modifies shared data inside a critical region, all update messages are buffered and delayed until the processor leaves the critical region Processes accessing shared data variables outside critical regions do it at their own risks –Same as with shared memory model –Risk is higher

21. FOUR CONSISTENCY PROTOCOLS 1.Conventional shared variables: –Replicated on demand – Single writer/multiple readers policy uses an invalidation-based protocol 2.Read-only variables: –Replicated on demand –Any attempt to modify them will result in a runtime error

22. FOUR CONSISTENCY PROTOCOLS 3.Migratory variables: –Migrated among the processes accessing them –Every process accessing them will always get full read and write access 4.Write-shared variables: –Can be updated concurrently because different portions of the page are accessed

23. Implementation Programmer uses annotations to specify any of the last three consistency protocols –Read-only variables –Migratory variables –Write-shared variables Incorrect annotations may result in inefficient performance or in runtime errors but not in incorrect results

24. WRITE-SHARED PROTOCOL (I) Designed to fight false sharing Uses a copy-on-write mechanism Whenever a process is granted access to write- shared data, the page containing these data is marked copy-on-write First attempt to modify the contents of the page will result in the creation of a copy of the page modified (the twin ).

25. x = 1 y = 2 x = 1 y = 2 First write access twin x = 3 y = 2 Before After Compare with twin New value of x is 3 Example

26. WRITE-SHARED PROTOCOL (II) At release time, the DSM will perform a word by word comparison of the page and its twin, store the diff in the space used by the twin page and notify all processors having a copy of the shared data of the update A runtime switch can be set to check for conflicting updates to write-shared data.

27. UPDATE TIME-OUT MECHANISM Munin does not send updates to processors holding stale replicas Anytime a processor receives an update for a page for which it does not have a twin, the page is marked supervisor-only and the time of receipt of the update is recorded. First local access to the page will cause a trap that will remove the restriction

28. UPDATE TIME-OUT MECHANISM When a process receives an update for a page that is still marked supervisor only, it checks the timestamp of the last update If more than 50 ms have elapsed, process notifies the originator of the update not to send more updates and invalidates the page.

29. CONCLUSIONS (I) The strongest point of Munin is its excellent performance –typically within 5 to 33% of the performances of hand-coded message passing versions of the same programs Its major limitation is its dependence of some features of the V kernel

30. CONCLUSIONS (II) Munin requires programs to access shared data from within critical regions or after barriers –Appears to be a reasonable requirement Munin allows users to tune the performance of their programs by selecting the best consistency protocol for each shared variable –Can quickly become a tedious process

31. FURTHER DEVELOPMENTS Same team has come with a successor to Munin named TreadMarks Key differences are: –TreadMarks uses a more complex lazy release protocol –TreadMarks is UNIX-based More portable