I joined Google in Sep '17. As a PhD student I worked with the Architecture Research Group at Simon Fraser University advised by Arrvindh Shriraman. I was also supervised by Nick Sumner, Viji Srinivasan and Sasha Fedorova.
My research can be broadly described as generalized methods for application specific hardware specialization.
I have worked on cache memory systems, coherence protocols, workload characterization and application specific hardware
specialization.
Publications : HPCA'18, HPCA'17, IISWC'16, MICRO'16, ICS'16, ISCA'15, ICS'15, ISCA'13, MICRO'12
Resume:
The need for computational power has increased with emerging workloads such as
machine learning driving the information economy.
Yet data center and embedded computing alike
desire reduced power consumption.
With the breakdown of semiconductor technology scaling, we no longer attain improved
performance and reduced energy consumption by using smaller transistors.
The semiconductor industry has
resorted to hardware specialization as a means to provide the
performance and energy benefits consumers have experienced since the
invention of the microprocessor. However, specialized hardware units
introduce new challenges, namely -- what to specialize,
how to specialize and how to integrate specialized units. Current approaches require
manual effort to analyse, restructure and rewrite workloads to take advantage
of specialized hardware accelerators.
My research has focused on automated compiler techniques with a goal
of specialization.
It is the first to directly address the problems of
what to specialize and how to specialize program
regions in large, irregular workloads.
I recognized the utility of adapting program analysis
techniques to address the problem.
My work at IISWC'16 demonstrates that analyses at coarser granularities
smears characteristics critical to hardware specialization.
Using a methodology adapted from program analysis,
I have synthesized an accelerator workload suite to assist researchers.
I have also developed program abstractions to ease the mapping of irregular
programs onto specialized hardware accelerators (HPCA'17).
The tools I have built are available as free and
open source software. Built upon a modern compiler framework,
they allow researchers to easily reproduce and extend the work that I have done.
In addition to this line of work, I have also conducted research on mitigating
energy consumption due to extraneous data movement.
In our work presented at ISCA'15, we showed the utility of time-stamp based coherence
protocols to reduce the overhead of data movement in accelerator rich architectures.
It allows for close integration of multiple specialized units
while easing the programming burden of explicit data transfers.
Furthermore, I have designed and evaluated adaptive granularity caching mechanisms to
eliminate waste (MICRO'12). This technique also provides a convenient substrate
for subsequent optimizations such as cache compression.
Contact:
TASC1 8208,
Simon Fraser University, BC,
Canada V5A1S6
NACHOS: Software-Driven Hardware-Assisted Memory Disambiguation for Accelerators
24th ACM International Conference on High Performance Computer Architecture
( Acceptance rate = 20% )
2018
[paper]
Needle: Leveraging Program Analysis to Extract Accelerators from Whole Programs
23rd ACM International Conference on High Performance Computer Architecture
( Acceptance rate = 22% )
2017
[paper]
Peruse and Profit: Estimating the Accelerability of Loops
30th ACM International Conference on Supercomputing
( Acceptance rate = 24% )
2016
[paper]
SPEC-AX: Extracting Accelerator Benchmarks from Microprocessor Benchmarks
2016 IEEE International Symposium on Workload Characterization
( Acceptance rate = 30% )
2016
[paper]
ChainSaw: Creating Von-Neumann Accelerators with Fused Instruction Chains
49th Annual IEEE/ACM International Symposium on Microarchitecture
( Acceptance rate = 22% )
2016
[paper]
Fusion: Design Tradeoffs in Coherent Cache Hierarchies for Accelerators
42nd Annual International Symposium on Computer Architecture
( Acceptance rate = 19% )
2015
[paper]
DASX: Hardware Accelerator for Software Data Structures
29th ACM International Conference on Supercomputing
( Acceptance rate = 25% )
2015
[paper]
Protozoa: Adaptive Granularity Cache Coherence
40th Annual International Symposium on Computer Architecture
( Acceptance rate = 19% )
2013
[paper]
Amoeba-Cache: Adaptive Blocks for Eliminating Waste in the Memory Hierarchy
45th Annual IEEE/ACM International Symposium on Microarchitecture
( Acceptance rate = 18% )
2012
[paper]