Berkeley UPC

1
Berkeley UPC
Kathy Yelick Christian Bell, Dan Bonachea, Wei
Chen, Jason Duell, Paul Hargrove, Parry Husbands,
Costin Iancu, Rajesh Nishtala, Mike Welcome LBNL
and U.C. Berkeley http//upc.lbl.gov
2
Berkeley UPC Compiler Status

• Recent Berkeley UPC release (v2.2)
• Support 1.2 language spec
• Supports collectives (tuning ongoing) memory
  model compliance
• Supports UPC I/O (naïve reference implementation)
• Compiler work
• Optimization phase and improved performance in
  v2.2
• Work on automated communication overlap,
  upc_forall,
• Large effort in quality assurance and robustness
• Test suite 600 tests run nightly on 20
  platform configs
• gt30,000 UPC compilations and gt20,000 UPC test
  runs per night
• Test suite infrastructure extended to support any
  UPC compiler
• now running nightly with GCC/UPC UPCR
• also has been used on HP-UPC, Cray UPC

3
Berkeley UPC Collaborations

• GCC UPC on Berkeley UPC Runtime
• Use for cluster (GASNet) implementations
• Now works with pthread runtime
• Source-level debugging with Totalview 7.x
• Joint project with Etnus
• General framework for source-to-source
  translators
• Future work
• Cray XT3 and other Rainier/Adams port
• Possible BlueGene/L port
• XT3 and BG/L both run on MPI conduit

4
Berkeley Applications Benchmarks

• Some new applications
• FT .45 TFlops on 512 proc Itanium/Quadrics
  (Elan4)
• CG 30 GFlops on 512 HP Alpha/Quadrics (Elan3)
• LU gt2 TFlops on 512 proc Itanium/Quadrics
  (Elan4)
• Barnes-Hut fine-grained (based on Splash)
• CFG uses to Chombo
• More on LU
• Towards a Sparse direct solver (SuperLU)
• Currently a full (top500-compliant) HPL
  implementation
• All UPC except for call to the BLAS

5
End of Berkeley Status
6
Data Movement and Synchronization
7
Motivation for Data Movement Synchronization

• Some are (at best) hard/slow in UPC
• Benchmarks highlight these
• FT communication-limited, all-to-all want to
  overlap
• MG fill in ghost regions
• Remote writes are often faster than remote reads
• But need to synchronize let the other proc know
  data is available
• See Tarek and John Mellor-Crummeys PPoPP05 paper
• Signaling store in Split-C
• Implementation issue reordering
• LU remotely enqueue a task
• GUPS and Histogram remotely increment/XOR a
  value
• With or without atomicity

8
Who Would Like to Talk?

• Non-Blocking Memget/put (Dan)
• Semaphores (Dan)
• Semaphore example (Tarek)
• Remote Atomics (Phil)
• Floating functions (Jason)