Azul@Purdue

Under CNS grant 0958465 "A Computational Infrastructure for Scalable Transactional Memory Abstractions" from the National Science Foundation, PIs Jagannathan and Vitek are heading a project on programming languages and paradigms for highly multithreaded applications. The research will involve experiments utilizing virtualization technology provided by Azul Systems. The S3Lab has acquired two Azul Compute Appliances, a Vega 3 Series 3300 with 108 cores and a Vega 3 Series 7300 with 864 cores. Purdue is the first academic institution in the country to have on-site access to these systems. Azul has also provided numerous APIs to allow students the ability to develop customized compilers, applications, and runtime environments that target these appliances.

The hardware platform are two Azul Appliances built around the Vega 3 processor. Featuring:

• Vega Processor: first 54-core chip designed and optimized for Java workloads. Featuring read and write barriers for garbage collection and object relocation, speculative locking to enable safe concurrent execution of code, an instruction set designed for the needs of virtual machines, predictive failure analaysis for de-configuration, and system memory chip-kill technology.
• Pauseless GC: a hardware-assisted garbage collection isolating GC execution onto separate processor cores.
• Highly Scalable Virtual Machine: optimized for Java workloads, offering up to 864 processors and 768 GB of true 64 bit SMP capacity.
• Optimistic Thread Concurrency: overcomming concurrency bottlenecks without requiring developers to optimize existing code.
• Cooperative Memory Management: support applications with additional memory borrowed from a shared memory pool. preventing or delaying out of memory errors that would have crashed the application and gives system administrators time to take appropriate action.
• Realtime Performance Monitoring: a diagnostics and tuning tool that offers a uniquely detailed view of all performance aspects of applications, including thread, memory, and IO profiles with zero overhead.