Tools and Environments for Parallel and Distributed Computing

ion. The TreadMarks application programming interface (API) provides facilities for process creation and destruction, synchronization, and shared memory allocation. Synchronization, a way for the programmer to express ordering constraints between the shared memory accesses of different processes, is implemented with critical sections. TreadMarks provides two synchronization primitives: barriers and exclusive locks. Barriers are global in the sense that calling the barrier process is stalled until all the processes in the system have arrived at that barrier. In the case of locks, a lock-acquire call acquires a lock for the calling process and a lock-release call releases it. TreadMarks uses multiple-writer protocol. The shared page is initially write-protected. When a write occurs in a processor (say P1), TreadMarks creates a copy of the page, or a twin, and saves it as a part of TreadMarks’ data structure on P1. It then un-protects the page in the user’s address space so that further writes to that page occur without software intervention. Later, P1 arrives at a barrier; there is an unmodified twin and a modified copy in the user’s address space. By making a word-by-word comparison of the two, a runlength encoding of the modifications of the page, called a diff, is created. Once the diff is created, it is sent to all the processors sharing that page. These processors then modify the page, discarding the twin. The same sequence of events takes place on every other processor. Once the diff is received, the entire sequence of events is local to each processor and does not require message exchanges, unlike in single-writer protocols. Brazos Brazos [6] is a page-based DSM that makes use of relaxed consistency models and multithreading on a network of multiprocessor computers. It executes on x86 multiprocessor workstations running Windows NT 4.0. Brazos is based on selective multicast in a time-multiplexed network environment such as Ethernet. Selective multicast is used in Brazos to reduce the number of consistency-related messages and to efficiently implement its version of scope consistency. One disadvantage with multicast is the potential harmful effect of unused indirect diff (i.e., run-length encoding of the modifications of a page). Although receiving multicast diffs for inactive pages does not increase network traffic, it does cause processors to be interrupted frequently to process incoming multicast messages. These messages and subsequent changes are not accessed before the next time that page is invalidated; thus, they detract user-code computation time. The dynamic copyset reduction mechanism ameliorates this effect by allowing processes to drop out of the copyset for a particular page. This causes them to be excluded from multicast messages, providing diffs for the page. Brazos uses multithreading at both user level and DSM system level. Multiple user-level threads allow applications to take advantage of SMP servers by using all available processors for computation. The Brazos runtime system has two threads. One thread is responsible for responding quickly to asynchronous requests for data from other processes and runs at the highest CLASSIFICATION OF DISTRIBUTED SHARED MEMORY SYSTEMS 71