Exploiting fine-grain thread level parallelism on the MIT multi-ALU processor

Abstract

Modern computer systems extract parallelism from problems at two extremes of granularity: instruction-level parallelism (ILP) and coarse-thread parallelism. VLIW and superscalar processors exploit ILP with a grain size of a single instruction, while multiprocessors extract parallelism from coarse threads with a granularity of many thousands of instructions.The parallelism available at these two extremes is limited. The ILP in applications is restricted by control flow and data dependencies [17], and the hardware in superscalar designs is not scalable. Both the instruction scheduling logic and the register file of a superscalar grow quadratically as the number of execution units is increased. For multicomputers, there is limited coarse thread parallelism at small problem sizes and in many applications. 1The research described in this paper was supported by the Defense Advanced Research Projects Agency and monitored by the Air Force Electronic Systems Division under contract F19628-92-C-0045. This paper describes and evaluates the hardware mechanisms implemented in the MIT Multi-ALU Processor (MAP chip) for extracting fine-thread parallelism. Fine-threads close the parallelism gap between the single instruction granularity of ILP and the thousand instruction granularity of coarse threads by extracting parallelism with a granularity of 50-1000 instructions. This parallelism is orthogonal and complementary to coarse-thread parallelism and ILP. Programs can be accelerated using coarse threads to extract parallelism from outer loops and large co-routines, finethreads to extract parallelism from inner loops and small subcomputations, and ILP to extract parallelism from subexpressions. As they extract parallelism from different portions of a program, coarse-threads, fine-threads, and ILP work synergistically to provide multiplicative speedup.These three modes are also well matched to the architecture of modern multiprocessors. ILP is well suited to extracting parallelism across the execution units of a single processor. Finethreads are appropriate for execution across multiple processors at a single node of a parallel computer where the interaction latencies are on the order of a few cycles. Coarse-threads are appropriate for execution on different nodes of a multiprocessor where interaction latencies are inherently 100s of cycles.Low overhead mechanisms for communication and synchronization are required to exploit fine-grain thread level parallelism. The cost to initiate a task, pass it arguments, synchronize with its completion, and return results must be small compared to the work accomplished by the task. Such inter-thread interaction requires 100s of cycles (2 13 s) on conventional multiprocessors, and 1000s of cycles (2 103 s) on multicomputers. Because of these high overheads, most parallel applications use only coarse threads, with many thousands of instructions between interactions.The Multi-ALU Processor (MAP) chip provides three on-chip processors and methods for quickly communicating and synchronizing among ...

Implementation and Evaluation of a Dynamically Routed Processor Operand Network

“…The M-Machine employed an on-chip cluster switch to connect the register bypass networks for three processors; an instruction writing to a remote register injects its result into the switch, which delivers the data to a waiting instruction on a remote processor [6]. The MIT RAW processor took this strategy further, by using a 4x4 mesh network to interconnect its processor tiles between execution units [7].…”

On the Role of Deterministic Fine-Grain Data Synchronization for Scientific Applications: A Revisit in the Emerging Many-Core Era

“…Unlike the full/empty bits like fine-grain synchronization [21,3,11,2,17,15], which tags the entire memory of the machine by associating additional access state bits with each word in memory, the design of SSB is motivated by the following observation: at any instance only a small fraction of memory locations is actively participating in synchronization [25].…”

“…HEP [21], Tera [3], MDP [11], Alewife [18,2], MMachine [17], Cray MTA-2 [1], the MT processor in Eldorado [15], and others associate additional access state bits (e.g., full/empty bits) with each word in entire memory. Fine-grain synchronization is achieved by accessing those word-level state bits in memory.…”

Interconnect Scaling Implications for CAD

“…We applied their model to four specifically synthesized blocks: three units from the MMachine, a fine-grained multicomputer designed at MIT and Stanford [14], and the global placement of Magic, a synthesized controller chip from Stanford's Flash multiprocessor [15] (minus the artificially long hand-routed MiscBus). From Figures 4 and 5 (we only show one M-Machine plot for brevity), we see that the model is a good fit for the wire length distributions of these designs, which span a wide range of gate count.…”