JavaFlow : a Java DataFlow Machine

Abstract

The JavaFlow, a Java DataFlow Machine is a machine design concept implementing a Java Virtual Machine aimed at addressing technology roadmap issues along with the ability to effectively utilize and manage very large numbers of processing cores. Specific design challenges addressed include: design complexity through a common set of repeatable structures; low power by featuring unused circuits and ability to power off sections of the chip; clock propagation and wire limits by using locality to bring data to processing elements and a Globally Asynchronous Locally Synchronous (GALS) design; and reliability by allowing portions of the design to be bypassed in case of failures. A Data Flow Architecture is used with multiple heterogeneous networks to connect processing elements capable of executing a single Java ByteCode instruction. Whole methods are cached in this DataFlow fabric, and the networks plus distributed intelligence are used for their management and execution. A mesh network is used for the DataFlow transfers; two ordered networks are used for management and control flow mapping; and multiple high speed rings are used to access the storage subsystem and a controlling General Purpose Processor (GPP). Analysis of benchmarks demonstrates the potential for this design concept. The design process was initiated by analyzing SPEC JVM benchmarks which identified a small number methods contributing to a significant percentage of the overall ByteCode operations. Additional analysis established static instruction mixes to prioritize the types of processing elements used in the DataFlow Fabric. The overall objective of the machine is to provide multi-threading performance for Java Methods deployed to this DataFlow fabric. With advances in technology it is envisioned that from 1,000 to 10,000 cores/instructions could be deployed and managed using this structure. This size of DataFlow fabric would allow all the key methods from the SPEC benchmarks to be resident. A baseline configuration is defined with a compressed dataflow structure and then compared to multiple configurations of instruction assignments and clock relationships. Using a series of methods from the SPEC benchmark running independently, IPC (Instructions per Cycle) performance of the sparsely populated heterogeneous structure is 40% of the baseline. The average ratio of instructions to required nodes is 3.5. Innovative solutions to the loading and management of Java methods along with the translation from control flow to DataFlow structure are demonstrated.