Today at 4 PM in 1325 CS, Milo Martin is going to give his academic
practice interview talk. Please try to attend and give Milo
feedback. Thanks to the pl-seminar folks for allowing us to use their
room, and they are welcome to attend as well.
The title and abstract are attached.
-Alaa
Computer Architecture Seminar: Milo Martin, UW-Madison,
"Token Coherence: Enabling Faster Multiprocessor Servers by
Decoupling Performance and Correctness"
This seminar is a practice academic interview talk. I'd welcome and
appreciate any feedback on my talk, especially from non-architects.
Token Coherence is a new hardware technique for increasing the performance
of commercial server workloads (e.g., database and web serving) running on
moderate-sized multiprocessors. These shared-memory multiprocessor servers
use cache-coherence protocols to provide the abstraction of a unified shared
memory, and the performance of existing cache-coherence protocols is
constrained either by requiring global message ordering or by the extra
latency added by request indirections.
Token Coherence provides a framework for avoiding these performance
bottlenecks by decoupling correctness requirements from performance
optimizations. In this new framework, the correctness substrate provides
strong correctness invariants involving token counting that guarantee
correct behavior in all cases (without requiring ordering and
indirections). A separate performance protocol provides high
performance in the common case, relying on a more conservative (less
efficient) mechanism only in rare cases. This approach
(1) avoids request ordering, (2) reduces indirections,
(3) enables predictive and adaptive techniques that can further
increase performance and scalability, and (4) may reduce the
verification effort required to eliminate protocol design errors that
can cause data corruption.
|