[talks] Jordan Fix will present his Pre-FPO on Monday, November 7th, 2016 at 4pm in CS 401

[Nicki Gotsis]

Jordan Fix will present his Pre-FPO on Monday, November 7th, 2016 at 4pm in CS 401.

The members of his committee are: 
Advisor: David August
Non-readers: David Wentzlaff, Margaret Martonosi
Readers: Andrew Appel, David Callahan (Facebook)

Everyone is invited to attend his talk. 

Abstract:
Speculation with transactional memory (TM) systems helps programmers and
compilers produce profitable thread-level parallel programs. Prior work shows
that supporting transactions that span multiple threads, rather than requiring
transactions be contained within a single thread, enables new types of
speculative parallelization techniques for both programmers and parallelizing
compilers. The only prior TM systems with this multithreaded transactional (MTX)
support are software-based TM (STM) systems. Unfortunately, these systems are
notorious for their high overheads when validating speculative memory
operations, which often make otherwise good parallelization unprofitable. Thus,
parallelized programs using these prior STM systems with MTX support were forced
to overcome this problem through significant manual effort by the programmer to
optimize away most of this speculation validation. Unfortunately, compiler
technology and static analyses have not proven strong enough to relieve the
programmer of this burden.

This thesis presents the first complete design and implementation of a hardware
TM (HTM) system with MTX support. This system provides low overhead speculation
validation, enabling the aforementioned new types of speculative parallelization
techniques to achieve good performance even with high amounts of speculation
validation. Profitable parallelization of complex programs can be achieved even
with maximal speculation validation of every load and store inside transactions
of tens to hundreds of millions of instructions, thus making parallelization
less laborious and more feasible for both programmers and compilers.

In addition to supporting MTXs, transactions in this system are more resilient
than in many other prior works. Transactions avoid false misspeculation due to
branch misprediction; provide for large read and write sets that commit in a
lazy fashion; and allow for context switches (and thread migration) due to
common operations such as preemption or virtual memory management. This allows
the system to accommodate the large and long running transactions that are
needed to parallelize complex programs. Across 6 SPEC benchmarks a geomean
speedup of 2.11X over sequential execution is achieved on a multicore machine
with 4 cores, while increasing energy usage by 10%.