Grant: Discovery Grants Program - Individual
University of British Columbia
$22,000.00
Memory coherence for emerging datacentre applications
FY 2016-17
Summary
OBJECTIVES
In this project we propose to develop a family of latency-tolerant, bandwidth-sensitive, and footprint-limited memory coherence protocols. Unlike traditional protocols, which are designed for homogeneous systems with custom fast interconnects, our protocols will operate efficiently in today's complex heterogeneous systems where inter-component latencies can vary by two orders of magnitude.
We will investigate latency-tolerant coherence in two emerging datacentre contexts: fine-grained programmable accelerators and oblivious computation. Because coherence is key to the adoption of both, the proposed research has the potential for significant advancement in both areas.
APPLICATION DOMAINS
Programmable accelerators have already been deployed on an industrial scale in such datacentre applications as web search (e.g., by Microsoft), and Intel's recent acquisition of Altera suggests that fine-grained programmable fabrics could soon be included in commodity processors. Compared to commodity CPUs, however, accelerators are much more difficult to program; we see this as arising largely from lack of effective support for coherent shared memory, the de facto enabling technology of parallel programming. This difficulty has formed a significant barrier to adoption despite the potential for significant performance and energy advantages offered by accelerators: it is difficult to justify putting months into developing accelerators for web search (e.g., Microsoft Catapult) when the machine learning models used to rank search results change several times a week.
At the same time, enterprises with cloud-computing interests are increasingly concerned about security. Trace-oblivious computing allows a client to securely compute in the cloud with the provider unable to snoop on the computation via temporal or spatial access patterns, but does so only at the cost of an order-of-magnitude penalty to performance. To overcome this cost, oblivious computation must leverage parallelism, and this, in turn, requires a secure variant of coherent shared memory.
The coherence techniques we propose will (a) make emerging accelerator fabrics far easier to program and dramatically reduce design cycles; and (b) enable parallelism in oblivious computing and allow its adoption in datacentres.
IMPACT
Datacentres are everywhere: even an enormous cloud provider like Google accounts for less than 1 of total datacentre computing globally, with most datacentres widely distributed in diverse industrial, governmental, and academic settings in Canada and abroad. The proposed research will have lasting impact on all of them: firstly, by speeding adoption of energy-efficient heterogeneous computing fabrics, it will allow datacentres to scale existing computing products; secondly, by enabling secure computing in the cloud, it will allow datacentres to expand to new application domains.
Program
Discovery Grants Program - Individual
Recipient
Location
British Columbia, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data
As published
Application id
RGPIN-2016-04265
Name nom
Lis, Mieszko
Department département
Electrical and Computer Engineering - Electrical and Computer Engineering
Organization id
2
Institution établissement
University of British Columbia
Province en
British Columbia
Province fr
Colombie-Britannique
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2016
Competition year année de concours
2016
Award amount
22000
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
Discovery grants programs
Group fr
Programmes de subventions à la découverte
Committee code
1507
Committee name en
Computer Science
Committee name fr
Sciences informatiques
Area of application code
801
Area of application group en
Information and communication services
Area of application group fr
Services d'information et de télécommunications
Area of application en
Computer software
Area of application fr
Logiciel
Research subject code
2702
Research subject group en
Information technology
Research subject group fr
Technologie de l'information
Research subject en
Computer systems organization
Research subject fr
Organisation des systèmes informatiques
Partie
1
Nb partie
5
Application title
Memory coherence for emerging datacentre applications
Keyword
NSERC
Application summary
OBJECTIVES
In this project we propose to develop a family of latency-tolerant, bandwidth-sensitive, and footprint-limited memory coherence protocols. Unlike traditional protocols, which are designed for homogeneous systems with custom fast interconnects, our protocols will operate efficiently in today's complex heterogeneous systems where inter-component latencies can vary by two orders of magnitude.
We will investigate latency-tolerant coherence in two emerging datacentre contexts: fine-grained programmable accelerators and oblivious computation. Because coherence is key to the adoption of both, the proposed research has the potential for significant advancement in both areas.
APPLICATION DOMAINS
Programmable accelerators have already been deployed on an industrial scale in such datacentre applications as web search (e.g., by Microsoft), and Intel's recent acquisition of Altera suggests that fine-grained programmable fabrics could soon be included in commodity processors. Compared to commodity CPUs, however, accelerators are much more difficult to program; we see this as arising largely from lack of effective support for coherent shared memory, the de facto enabling technology of parallel programming. This difficulty has formed a significant barrier to adoption despite the potential for significant performance and energy advantages offered by accelerators: it is difficult to justify putting months into developing accelerators for web search (e.g., Microsoft Catapult) when the machine learning models used to rank search results change several times a week.
At the same time, enterprises with cloud-computing interests are increasingly concerned about security. Trace-oblivious computing allows a client to securely compute in the cloud with the provider unable to snoop on the computation via temporal or spatial access patterns, but does so only at the cost of an order-of-magnitude penalty to performance. To overcome this cost, oblivious computation must leverage parallelism, and this, in turn, requires a secure variant of coherent shared memory.
The coherence techniques we propose will (a) make emerging accelerator fabrics far easier to program and dramatically reduce design cycles; and (b) enable parallelism in oblivious computing and allow its adoption in datacentres.
IMPACT
Datacentres are everywhere: even an enormous cloud provider like Google accounts for less than 1 of total datacentre computing globally, with most datacentres widely distributed in diverse industrial, governmental, and academic settings in Canada and abroad. The proposed research will have lasting impact on all of them: firstly, by speeding adoption of energy-efficient heterogeneous computing fabrics, it will allow datacentres to scale existing computing products; secondly, by enabling secure computing in the cloud, it will allow datacentres to expand to new application domains.