Tasuku Hiraishi

Lightweight lexical closures for legitimate execution stack access

We propose a new language concept called “L-closures” for a running program to legitimately inspect/modify the contents of its execution stack. L-closures are lightweight lexical closures created by evaluating nested function definitions. A lexical closure can access the lexically-scoped variables in the creation-time environment and indirect calls to it provide legitimate stack access. By using an intermediate language extended with L-closures in high-level compilers, high-level services such as garbage collection, check-pointing, multithreading and load balancing can be implemented elegantly and efficiently. Each variable accessed by an L-closure uses private and shared locations for giving the private location a chance to get a register. Operations to keep coherency with shared locations as well as operations to initialize L-closures are delayed until an L-closure is actually invoked. Because most high-level services create L-closures very frequently but call them infrequently (e.g., to scan roots in garbage collection), the total overhead can be reduced significantly. Since the GNU C compiler provides nested functions, we enhanced GCC at relatively low implementation costs. The results of performance measurements exhibit quite low costs of creating and maintaining L-closures.

Automatic Parameter Tuning of Three-Dimensional Tiled FDTD Kernel

This paper introduces an automatic tuning method for the tiling parameters required in an implementation of the three-dimensional FDTD method based on time-space tiling. In this tuning process, an appropriate range for the tile size is first determined by trial experiments using cubic tiles. The tile shape is then optimized by using the Monte Carlo method. The tiled FDTD kernel was multi-threaded and its performance with the tuned parameters was evaluated on multi-core processors. When compared with a naively implemented kernel, the performance of the tuned FDTD kernel was improved by more than a factor of two.

Parallel Graph Traversals using Work-Stealing Frameworks for Many-core Platforms

Parallel programming/execution frameworks for many/multi-core platforms should support as many applications as possible. In general, work-stealing frameworks provide efficient load balancing even for irregular parallel applications. Unfortunately, naive parallel programs which traverse graph-based data structures (e.g., for constructing spanning trees) cause stack overflow or unacceptable load imbalance. In this study, we develop parallel programs to perform probabilistically balanced divide-and-conquer graph traversals. We propose a programming technique for accumulating overflowed calls for the next iteration of repeated parallel stages. In an emerging backtracking-based work-stealing framework called “Tascell, ” which features on-demand concurrency, we propose a programming technique for long-term exclusive use of workspaces, leading to a similar technique also in the Cilk framework.

Reducing Redundant Search in Parallel Graph Mining Using Exceptions

This paper proposes an implementation of a parallel graph mining algorithm using a task-parallel language having exception handling features. The performance of a straightforward task-parallel implementation is poor for many practical backtrack search algorithms due to pruning employed in them, a worker may prune a useless subtree, which is pruned before traversal in the sequential search algorithms for search space reduction, after another worker starts the traversal of it resulting in a large amount of redundant search. Such redundancy will be significantly reduced by letting a worker know that the subtree which it is traversing is pruned so that it aborts the traversal. This abortion can be implemented elegantly and efficiently using the task-parallel language that has a mechanism for exception handling by which all running parallel tasks in a try block with an exception are automatically aborted. We applied this abort mechanism to the graph mining algorithm called COPINE, which is practically used for drug discovery, using the task-parallel language Tascell. As a result, we reduced the search space by 31.9% and the execution time by 27.4% in a 28-worker execution.

Extending a Work-Stealing Framework with Probabilistic Guards

We propose probabilistic guards and analyze their performance. To reduce the total task division cost, probabilistic guards can prevent thief workers from stealing small tasks from victim workers probabilistically. In this study, we have implemented probabilistic guards on a work-stealing framework called Tascell and have confirmed that they perform well. In theory, a thief may repeat an unbounded number of probabilistically prevented steal attempts until success if a victim uses a probabilistic guard that rejects steal attempts with a non-zero probability. Therefore, in this paper, we also propose a mechanism that invalidates probabilistic guards on demand by setting an upper limit to the number of repeated probabilistically prevented steal attempts. We evaluate its potential effects on probabilistic guards by measuring the actual numbers of repeated attempts until success. We also evaluate the performance of probabilistic guards with various upper limits. Finally, we propose and evaluate virtual probabilistic guards that act as probabilistic guards without repeating probabilistically prevented steal attempts and they exhibit superior performance.

An Implementation of Exception Handling with Collateral Task Abortion

In this paper, an implementation of exception handling for task-parallel languages is proposed such that all running parallel tasks in a try block with an exception are automatically aborted as soon as possible. In parallel tree search, exception handling that allows such a collateral task abortion is useful when the objective is to complete the search as soon as one solution is found or to allow a worker to abort the traversal of a subtree that is found to be redundant by another worker, even when it has been initiated. However, few existing task-parallel languages, such as Cilk Plus and X10, have this capability. In this study, we enhanced a task-parallel language, Tascell, with this capability. Since the Tascell compiler is implemented as a translator to C code, techniques are required for implementing the non-local exit mechanism with cleanup code execution in the “finally” clauses. We achieved this implementation by exploiting nested functions, which are already used in the temporary backtracking mechanism of Tascell. We also modified the task scheduler provided by Tascell such that a worker can abort a task after it is started. When aborting a task, the scheduler also aborts all its descendant tasks. We evaluated our implementation in terms of overheads and time taken to abort tasks.

Parallelization of Extracting Connected Subgraphs with Common Itemsets in Distributed Memory Environments

This paper proposes a parallel implementation of graph mining that extracts all connected subgraphs with common itemsets, of which the size is not less than a given threshold, from a graph and from itemsets associated with vertices of the graph, in distributed memory environments using the task-parallel language Tascell. With regard to this problem, we have already proposed parallelization of a backtrack search algorithm named COPINE and its implementation in shared memory environments. In this implementation, all workers share a single table, which is controlled by locks, that contains the knowledge acquired during the search to obviate the need for unnecessary searching. This sharing method is not practical in distributed memory environments because it would lead to a drastic increase in the cost of internode communications. Therefore, we implemented a sharing method in which each computing node has a table and sends its updates to the other nodes at regular time intervals. In addition to this, the high task creation cost for COPINE is problematic and thus the conventional work-stealing strategy in Tascell, which aims to minimize the number of internode work-steals, significantly degrades the performance since it increases the number of intranode work-steals for small tasks. We solved this problem by promoting workers to enable them to request tasks from external nodes. We also employed a work-stealing strategy based on estimation of the sizes of tasks created by victim workers. This approach enabled us to achieve good speedup performance with up to 8 nodes × 16 workers.

Evaluation of an MPI-Based Implementation of the Tascell Task-Parallel Language on Massively Parallel Systems

Tascell is a task parallel language that supports distributed memory environments. The conventional implementation of Tascell realizes inter-node communication with TCP/IP communication via Tascell servers. This implementation is suitable for dynamic addition of computation nodes and wide-area distributed environments. On the other hand, in supercomputer environments, TCP/IP may not be available for inter-node communication and there may be no appropriate places for deploying Tascell servers. In this study, we have developed a server-less implementation of Tascell that realizes inter-node communication with MPI communication in order to evaluate its performance on massively parallel systems. It performs well on four Xeon Phi coprocessors (with 456 workers) and the K computer, for instance, our 19-queens solver achieves a 4615-fold speedup relative to a serial implementation with 7168 workers on the K computer. Our server-less implementation realizes deadlock freedom, although it only requires the two-sided communication paradigm and the MPI_THREAD_FUNNELED support level. On Xeon Phi coprocessors, we compare our implementation with other implementations that employ TCP/IP or the MPI_THREAD_MULTIPLE support level.

Experience with SC: transformation-based implementation of various extensions to C

We have proposed the SC language system which facilitates language extensions by translation into C. In this paper, we present our experience with the SC language system and discuss its design, implementation, applications and improvements. In particular, we present the improvement to the design of transformation rules for implementing translations, which includes the feature to extend an existing transformation phase (rule-set). This enables us to implement many transformation rule-sets only by describing the difference, and helps us to use commonly-used rule-sets as part of the entire transformation. We also show several actual examples of extensions to C: garbage collection, multithreading and load balancing.