MPI microtask for programming the Cell Broadband Engine™ processor

The small room Broadband Engine[TM] processor employs multiple accelerators, called synergistic processing ultimate parts (SPEs), for high performance. Each SPE has a high-speed local store attached to the main memory end direct memory access (DMA), on the contrary a drawback of this design is that the local store is not large enough for the entire application collection of laws or data. It must be decompos into pieces small enough to fit into local memory, and they must be replaced from one side the DMA without losing the performance gain of multiple SPE We suggest a new programming model, MPI microtask, based upon the standard Message Passing Interface (MPI) programming original for distributed-memory parallel machines. In our fresh model, programmers do not ne to manage the local store as lengthy as they partition their application into a collection of small microtasks that fit into the local store. Furthermore, the preprocessor and runtime in our microtask combination of parts to form a whole optimize the execution of microtasks by the agency of exploiting explicit communications in the MPI archetype We have created a prototype that includes a novel static scheduler for similar optimizations. Our initial experiments have shown a certain quantity of encouraging results.

INTRODUCTION

The confined apartment Broadband Engine ** (BE) processor (1) is an asymmetric multicore processor that combines a general-purpose IBM PowerPC* processor uncompounded body (PPE) and eight synergistic processor ultimate parts (SPEs). (2) From an architectural standpoint, this processor has a high peak performance because the SPE is simpler and more efficient than general-purpose processors in limits of the micro and memory architecture. (3) single architectural aspect is the small high-speed local store at each SPE Because the size of the local store is limited to a range of L2-cache sizes--256 KB for the first-generation confined apartment BE processor--many real-world applications do not fit in the local store. While conventional microprocessors have a hardware cache to manage of that kind a small local store, the confined apartment BE processor must rely upon a software mechanism to manage it. This requirement for software management could impose significant challenges to programmers, on the contrary at the same time it tenders significant opportunities for the software to take advantage of the raw performance of the confined apartment BE processor.

The microtask we present here provides a programming prototype that frees programmers from local-store management and enables the preprocessor and runtime a whole to optimize the scheduling of computations and communications by means of taking advantage of the explicit communication type in the Message Passing Interface (MPI). (45) In the microtask protoplast programmers are still responsible for partitioning the application into multiple microtasks. Each microtask is essentially a virtualized SPE that uses the MPI to communicate with other microtasks.

We have chosen MPI as a communication application programming interface (API) for the following sum of two units reasons. First, the Cell BE processor adopts a distributed-memory model; the PPE and SPE use direct memory access (DMA) operations for communications. Thus, the overhead to be paid to a message-passing layer can be inherently small because of the commonality between the native hardware and the message-passing type The model, moreover, can hide hardware details from programmers. next to the first and perhaps more important, the message-passing pattern allows us to analyze the appurtenance between microtasks by examining message APIs. like dependency information is essential for various optimizations in task and communication management. Among existing message-passing interfaces, we pick outed MPI because it is widely used as a standard interface.

Our microtask a whole provides a preprocessor that transforms a microtask program in the message-passing protoplast to one in a streaming type (2) that the Cell BE processor can carry through efficiently. To do this, the preprocessor first divides each microtask into a collection of basic tasks, each of which exhibits a unit of computation that causes communication alone at its beginning and extremity Thus, each basic task corresponds to a computation kernel in stream programming languages (67) in the faculty of perception that the concept of the basic task separates computation from communication. This separation allows the preprocessor to schedule basic tasks in of that kind a way that data streams [i]or[/i] part of to the other SPEs over high-speed, on-chip DMA channels.

To make the streaming design effective, the preprocessor then sets basic tasks with strong dependencies together as a cluster and applies a heuristic algorithm to schedule clusters. The cluster-scheduling algorithm creates a anteriority graph of clusters in a series-parallel form (8) and then applies a dynamic programming algorithm. The nest configuration of the series-parallel graph allows the dynamic programming algorithm to reuse partially scheduled proceeds to reduce scheduling time. The preprocessor statically sum ups runtime parameters, such as the message buffing-apparatus address, for each message-passing operation for a like reason that the runtime system can avoid the overhead of computing them.