Citation Context

...ect of a shared region. An alternate approach is to use static type checking that distinguishes private and shared objects and prevents access to private objects from other threads. See Liblit’s work =-=[38]-=- for details on one such type system. Shared Region with Global Creation and Deletion For this type of region, the code executed by each thread must include an explicit call to an operation to create ...

Citation Context

...ally the major data structures reside in shared space). In Titanium, by contrast, all heap and static data can potentially be accessed remotely [13], although a sophisticated compiler escape analysis =-=[17] can detec-=-t which objects are potentially "shared" (interesting applications typically have about 50-100% of the total bytes allocated judged to be shared by the analysis). Dynamic allocations of shar...

Citation Context

...sharing reference qualifiers - Titanium’s type system is augmented with reference qualifiers that enable the expression and inference of locality and sharing properties of distributed data structures =-=[29, 30]-=-. 4. Explicitly unordered loop iteration - Titanium’s foreach loop construct provides a concise means for expressing iteration over an N-dimensional region or array domain and includes an explicitly u...

Citation Context

...e the number of fences needed to guarantee SC. These analyses leverage a number of techniques to determine whether a memory location can potentially be involved in a race, including sharing inference =-=[26]-=-, pointer alias analysis, and thread escape analysis. These techniques require fairly-complex whole-program analyses that are difficult to scale to large programs, especially for languages like C++. I...

Citation Context

...but the fence-insertion algorithm are interprocedural analyses that are fairly complex. Overall, their method induces slowdowns of over 10% on average over the relaxed Java Memory Model. Liblit et al =-=[17]-=- developed an SC version of Titanium [13]. In the same project, Krishnamurthy and Yelick [14] showed how the regular structure of SPMD programs could be exploited to reduce the complexity of delay set...

Citation Context

...ace. The Titanium type system does not separate levels 1 and 2 of the hierarchy. The distinction between 1 and 2 is important for many applications, such as race detection [21], data sharing analysis =-=[17]-=-, and sequential consistency enforcement [14], since references to level 1 values on different threads cannot be to the same location. Other applications such as data locality inference [16] can benef...

Citation Context

...global address space dialect of Java that runs on most parallel and distributed memory machines. The analyses can be used by many clients, including locality inference [33] and data sharing inference =-=[34]-=-. In this report, we consider two clients in particular. We first use them to perform data race analysis [36], which can be used to report potential program errors to application programmers. We then ...

Citation Context

...that are meaningful only to a single processor, and global pointers that are valid everywhere. The type system ensures that local pointers do not leak into a global context. This work was extended in =-=[13]-=- to add types for dealing with private vs. shared data. However, this line of work does not consider security properties that require defense against an adversary. Riely and Hennessey study type safet...

Citation Context

... data-sharing inference in the Titanium compiler using pointer analysis. The precision of this inference compared to Liblit, Aiken, and Yelick’s constraint-based sharing qualification inference (SQI) =-=[66]-=- is shown in Figure 8.35. As with locality inference, results are only reported for reachable local variables, fields, and method parameters and return values of reference type in the application code...

Citation Context

...ion with global collective communication and synchronization operations. The SPMD model’s simplicity enables a variety of program analyses, including data locality analysis [9], data sharing analysis =-=[10]-=-, concurrency analysis [6], and pointer analysis [7]. However, the flat SPMD model does not fit well with divide-and-conquer parallelism or hierarchical machines that mix shared and distributed memory...