Citation Context

...between the two environments. Finally, DSM allows for seamless integration of shared-memory multiprocessor workstations in a network environment. The actual system described in this paper, TreadMarks =-=[7]-=-, runs at user-level on Unix workstations. No kernel modifications or special privileges are required, and standard Unix compilers and linkers are used. The challenge in implementing a DSM system is t...

Citation Context

... the machine is either up or down. Shared state: Unlike SMPs, clusters have no shared state. Although much work has been done to emulate global shared state through software distributed shared memory =-=[31,32,34]-=-, we can improve performance and reduce complexity if we can avoid or minimize the need for shared state across the cluster. These two concerns, partial failure and shared state, lead us to focus on t...

Citation Context

...otations or explicit calls to access shared data [2, 11]. Another approach, called Shared Virtual Memory (SVM), uses the virtual memory hardware to detect access to data that is not available locally =-=[4, 13, 12]-=-. In most such systems, the granularity at which data is accessed and kept coherent is large, because it is related to the size of an application data structure or the size of a virtual page. We have ...

Citation Context

...hreaded processors [23], single-chip multiprocessors [16, 21], more traditional shared-memory multiprocessors of any size [14], and even multiprocessors built using software distributed shared-memory =-=[11]-=-. Our approach scales (both up and down) to all of these architectures because it is built upon writeback invalidation-based cache coherence, which itself scales to any of these machines. Our unified ...

Citation Context

... the parallel "minicomputers" of the future [1]. Unfortunately, current systems directly support only message-passing communication. Shared memory is limited to page-based systems, such as T=-=readMarks [17]-=-, which are not sequentially consistent and which can perform poorly in the presence of fine-grain data sharing [11]. These systems lack fine-grain access control , a key feature of hardware shared-me...

Citation Context

...h no additional processing. Therefore, by reducing the number of page invalidations the update cost is also proportionally reduced. A diff-based all software LRC scheme like the one used in TreadMarks=-=[13]-=- doesn't have the same property of keeping a home copy up-to-date. Updates are embedded in diffs which have to be transferred eventually and applied on every local copy. Therefore, postponing some inv...

Citation Context

...s on release, whereas the LRC protocol postpones update propagations until the next acquire. Although previous prototypes have shown reasonably good performance for some applications on small systems =-=[21]-=-, protocol overhead becomes substantial on large-scale systems. Our own experience shows that many applications do not speedup well using standard LRC-based shared virtual memory on a 32-node machine,...

Citation Context

...g node must record the identity of the blocks selected for self-invalidation and invalidate them at a point in the future that maximizes performance. Systems that maintain cache coherence in software =-=[24,26,33,35]-=- can use arbitrary data structures to store block identities. The blocks are self-invalidated using the same primitives required to process explicit invalidation messages. Alternatively, hardware mana...

Citation Context

...s Bytecode. By shipping Bytecode, a single resulting Bytecode file is sufficient instead of ten. "Java/DSM" [22] is an implementation of Java on top of the Treadmarks distributed shared memo=-=ry system [12]-=-. Java/DSM requires special implementations of the Java virtual machine since objects must be allocated on the heap. In addition, Java/DSM has to struggle with heterogeneous environments which are alr...