Dan Alistarh

Date: 17:00, Tuesday, February 16, 2016
Speaker: Dan Alistarh
Venue: IST Austria
Notes:

Day: Tuesday

Abstract:

Concurrent memory reclamation, that is, the problem of reclaiming
memory in a multi-threaded environment, is one of the main problems in
concurrent data structure design. Almost all known solutions induce
high overhead, or must be customized to the specific data structure by
the programmer.

In this talk, I will talk about two new solutions to this problem. The
first is StackTrack, a memory reclamation algorithm which can be
applied automatically by a compiler, and still be efficient. The key
idea behind StackTrack is to eliminate most of the bookkeeping
overheads by using the hardware transactional memory (HTM) extensions
offered by new Intel processors.
The second solution, called ThreadScan, is an extension of the
framework which removes the need for hardware TM support by relaxing
progress guarantees, and exploiting the operating system signaling
mechanisms. Empirical results show that both solutions match or
improve upon the performance of previous, non-automatic solutions for
memory reclamation.
My talk will also touch upon some technical and algorithmic issues
arising when using HTM for designing fast concurrent data structures.

This talk is based on joint work with Patrick Eugster, Maurice
Herlihy, Will Leiserson, Alexander Matveev, and Nir Shavit, which
appeared in EuroSys 2014 and SPAA 2015.

Short Bio:

Dan Alistarh is currently a researcher at Microsoft Research
Cambridge. Before MSR, Dan was a post-doc at MIT, working with Nir
Shavit and Nancy Lynch. He obtained his PhD from EPFL, under the
guidance of Rachid Guerraoui. Dan’s research focuses on building
scalable algorithms, and combines algorithmic ideas with practical
implementations.

Posted in RiSE Seminar