| Research
Compiler-Enabled Memory Systems
Traditional memory systems suffer from a certain degree of
redundancy because of the general one-size-fits-all approach (e.g.
doing a TLB lookup followed by associative cache lookup for each
memory access is wasteful both energy- and performance-wise). We
propose and evaluate a compiler-enabled memory system framework
that leverages static program information (i.e.,
compile-time information) about memory access types and patterns, to reduce some of the
aforementioned redundancy. We have used the SUIF/MachSUIF compiler
infrastructure for doing the compiler analyses. For performance
and energy evaluation, we run the binaries generated by MachSUIF
on the execution-driven SimpleScalar/Wattch simulator, which
models an Alpha processor with our proposed extensions. This
research has so far resulted in publications in premier conferences
on Architecture (IEEE Micro 2001 and ACM ASPLOS 2002; see below).
Fine-Grain Synchronization for Multiprocessors
Ever-increasing multiplicity in modern multiprocessor systems
has prompted researchers to explore new ways for making programs
highly concurrent to better exploit the offered capability. Efficient support for synchronization is set
to become a key challenge to allow scalability. Performance results of machines
supporting fine-grain synchronization, like the Alewife
multiprocessor, have demonstrated the benefits of fine-grain synchronization over
conventional coarse-grain synchronization.
The Alewife machine implements fine-grain synchronization as a
separate layer above the cache-coherence protocol. We on the other
hand, propose integrating fine-grain synchronization support in
the cache-coherence layer itself. We believe that this streamlined
solution can further increase performance in addition to reducing
network traffic (because we eliminate the inter-layer messages
required if fine-grain synchronization and cache-coherence were
implemented as two separate layers). To evaluate this proposal, we
are currently extending the SimpleScalar simulator to support
multiprocessing/multithreading.
Publications
Raksit Ashok, Saurabh Chheda, Csaba Andras Moritz (.ps)
(.pdf)
Cool-Mem: Combining Statically Speculative and Conventional Memory Accessing
for Energy Efficiency, To Appear in the International Conference on Architectural
Support for Programming Languages and Operating Systems (ASPLOS 2002).
Osman S. Unsal, Raksit Ashok, Israel Koren, C. Mani Krishna,
Csaba Andras Moritz (.ps) (.pdf)
Cool-Cache for Hot Multimedia, Appears in the IEEE International Symposium on
Microarchitecure (MICRO-34), Dec.2001
Raksit Ashok, Osman S. Unsal, Vijaya R. Lakamraju, Csaba Andras Moritz
(.ps) (.pdf)
Network Software: From NCP To Ubiquitious Computing, To appear in the
Encyclopedia Of Life Support Systems, 2001
Raksit Ashok, Diganta Roychowdhury, Csaba Andras Moritz
(.ps) (.pdf)
Synchronization Coherence Protocols: Unifying Synchronization and Caching
in Multiprocessors, UMass Technical Report.
|