FINAL LECTURE

Reports:
Intro
Serial Problem
Concurrent Architecture
Performance Results
Lessons Learned
Conclusion

What was this class about, really?
- something introductory that I missed
- not everyone is convinced that concurrency and parallelism are
  things that we're all going to be doing
- Intel sells sand, it's really cool sand, and they're really good at
  making it with low error rates

At Intel, Job 1 is sell processors. Job 1A is to encourage the
development of software that requires new processors. Microsoft makes
money through NEW COMPUTER SALES (no one upgrades). Moore's law is
kaput, and concurrency seems to be the only way to get performance
through the roof again (unless some new physical technology catches
on).

We looked at the nature of decomposition:
1) task parallelism (servers, deferring work, prefetching, speculation)
    - Challenge: eliminate dependencies
    - Pthreads, OpenMP, AME
2) data parallelism (big numbers, large structures)
   - TBB, OpenMP (goes both ways)

Dependency management boils down to constraints on the possible
execution paths of the program.

2 Flavors:
1) Conservative (prevention)
  - Mutexes & CVs
  - Atomic ops
2) Optimistic
  - Cache/verify -- handcoded, lock free, need to worry about memory model
  - Transactional memory


Programming Model

X Composibility
X Cache and verify is hard
X Transactional memory is hard