Compression Scheme of Log-Based Architectures
Tuesday February 19, 2008
Hamerschlag Hall D-210
4:00 pm
Theodoros Strigkos
Carnegie Mellon University
Runtime monitoring tools are invaluable for detecting various types of bugs, yet
rarely used for deployed code as they impose significant overheads on the monitored
application. Log-Based Architectures is an architectural approach that advocates
utilizing otherwise idle cores in a CMP to dramatically reduce the overheads for
runtime program monitoring. In order to achieve this, the main program’s trace
is logged and delivered to the monitoring core for inspection; however, the bandwidth
required for the log transport easily exceeds the bandwidth offered by available
on-chip interconnects.
In this talk, I will present the value-prediction based compression scheme of
Log-Based Architectures that reduces the log-related bandwidth and storage
requirements by an order of magnitude. The main innovation over previous
value-prediction compressors is the use of mechanisms that allow for nearly random
access in the compressed stream as well as the introduction of techniques such as
half prediction and default stride prediction that reduce the misprediction rate and
penalty, resulting to a compressed stream that requires less than 1 byte per
instruction.
Theodoros Strigkos is a Ph.D. student advised by Professor Babak Falsafi and
Professor Todd Mowry in Electrical and Computer Engineering at Carnegie Mellon
University. He received his Diploma in Engineering from the National Technical
University of Athens, Greece. He is mainly interested in the area of computer
architecture.
|