Efficient Parallel Programming on Datacenter Platforms with Disciplined Memory Access

MPhil Thesis Defence


Title: "Efficient Parallel Programming on Datacenter Platforms with
Disciplined Memory Access"

By

Mr. Xiang GAO


Abstract

Many distributed system technologies and programming frameworks are developed 
to support parallelized computation on the datacenter platform. While 
simplifying data-centric parallel computation, existing system technologies and 
application frameworks, however, exhibit limitation in their capability, 
efficacy and programmability in supporting sophisticated applications. In this 
paper, we present a distributed multitasking engine, together with a new 
programming language and a compiler, to support large-scale parallel computing 
on compute clusters with loosely-coupled commodity servers. The system, called 
Disciplined Runtime, is built under the principle of a disciplined memory 
access model, which means the programmer and the compiler furnish memory access 
information to the system so that the latter can detect memory access patterns 
and locality preferences at run time. Without limiting the expressing 
capability, this makes the memory access behavior known to the memory subsystem 
and the scheduler, and this information can be utilized to significantly 
improve the efficiency and performance of the system. The Disciplined Runtime 
provides distributed multitasking in a large uniform memory space with 
transactional semantics, and the new programming language enables programmers 
to provide memory access information of parallel tasks in a straightforward way 
so that the programs naturally follow the disciplined memory access model. 
Evaluation on both our research testbed and Amazon EC2 show that the new 
multitasking engine provides high performance and scalability on 32 physical 
compute servers and 256 large EC2 instances.


Date:			Friday, 21 September 2012

Time:			2:00pm – 4:00pm

Venue:			Room 3402
 			Lifts 17/18

Committee Members:	Dr. Lin Gu (Supervisor)
 			Dr. Charles Zhang (Supervisor)
 			Dr. Qiong Luo (Chairperson)
 			Prof. Shing-Chi Cheung


**** ALL are Welcome ****