What's New in 6.2

New Supported Platforms:

• POWER7 for BlueWaters development systems (new in 6.2.1)
• 64-bit MIPS, such as SiCortex, using mpi-linux-mips64
• Windows HPC cluster, using mpi-win32/mpi-win64
• Mac OSX 10.6, Snow Leopard (32-bit and 64-bit)

General Changes

• Runtime support
  • (new in 6.2.1) Topology information gathering has been optimized
  • (new in 6.2.1) Converse (seed) load balancers have many new optimizations applied
  • (new in 6.2.1) CPU affinity can be set more easily using +pemap and +commap options instead of the older +coremap
  • Smarter build/configure scripts
  • A new interface for model-based load balancing
  • new CPU topology API
  • a general implementation of CmiMemoryUsage()
  • Bug fix: Quiescence detection (QD) works with immediate messages
  • New reduction functions implemented in Converse
  • CCS (Converse Client-Server) can deliver message to more than one processor
  • Added a memory-aware adaptive scheduler, which can be optionally compiled in to charm
  • Added preliminary support for automatic message prioritization (disabled by default)
• Charm++
  • (new in 6.2.1) Better support for multicasts on groups
  • (new in 6.2.1) HybridLB (hierarchical balancing for very large core-count systems) has been substantially improved
  • (new in 6.2.1) Load balancing infrastructure has further optimizations and bug fixes
  • Cross-array and cross-group sections
  • Structured Dagger (SDAG): Support templated arguments properly
  • Plain chares support checkpoint/restart (both in-memory and disk-based)
  • Conditional packing of messages and parameters in SMP scenario
  • Changes to the CkArrayIndex class hierarchy
    • sizeof() all CkArrayIndex* classes is now the same
    • Codes using custom array indices have to use placement-new to construct their custom index. Refer example code: examples/charm++/hello/fancyarray/
    • Backward Incompatibility: CkArrayIndex[4D/5D/6D]::index are now of type int (instead of short) However the data is stored as shorts. Access by casting CkArrayIndexND::data() appropriately
    • Deprecation: The direct use of public data member CkArrayIndexND::index (N=1..6) is deprecated. We reserve the right to change/remove this variable in future releases of Charm++. Instead, please access the indices via member function: int CkArrayIndexND::data()
• Adaptive MPI (AMPI)
  • (new in 6.2.1) AMPI now implements MPI_Issend
  • (new in 6.2.1) AMPI's MPI_Alltoall uses a flooding algorithm more agressively, versus pairwise exchange
  • Compilers renamed to avoid collision with host MPI (ampicc, ampiCC, ampif77, ampif90)
  • Improved MPI standard conformance, and documentation of non-conformance:
    • Bug fixes in: MPI_Ssend, MPI_Cart_shift, MPI_Get_count
    • Support MPI_IN_PLACE in MPI_(All)Reduce
    • Define various missing constants
  • Return the received message's tag in response to a non-blocking wildcard receive, to support SuperLU
  • Improved tracing for BigSim
• Multiphase Shared Arrays (MSA)
  • Typed handles to enforce phases
  • Split-phase synchronization to enable message-driven execution
  • 3D arrays
• TCharm:
  • (new in 6.2.1) Object mappings can be read from a file, to allow offline topology-aware placement
  • Automatic tracing of API calls for simulation and analysis
• Debugging
  • Wider support for architectures other than net- (in particular MPI layers)
  • Improved support for large scale debugging (better scalability)
  • Enhanced record/replay stability to handle various events, and to signal unexpected messages
  • New detailed record/replay: The full content of messages can be recorded, and a single processor can be re-executed outside of the parallel application
• Performance analysis:
  • Tracing of nested entry methods
  • (new in 6.2.1) Projections logs can be spread across multiple directories, speeding up output when dealing with thousands of cores (+trace-subdirs N will divide log files evenly among N subdirectories of the trace root, named PROGNAME.projdir.K)
• Automatic Performance Tuning: Created an automatic tuning framework [still for experimental use only]
• CkMulticast: Network-topology / node aware spanning trees used internally for and lower bytes on the network and improved performance in multicasts and reductions delegated to this library
• Comlib: Improved OneTimeMulticastStrategy classes
• BigSim
  • Out-of-core support, with prefetching capability
  • Detailed tracing of MPI calls
  • Detailed record/replay support at emulation time, capable of replaying any emulated processor after obtained recorded logs.
• (new in 6.2.1) Virtualized ARMCI support has been extended to cover the functions needed by Coarray Fortran (CAF)

Architecture-specific changes

• Net-*: Can run jobs with more than 1024 PEs
• Net-* (new in 6.2.1): Can use clusters' mpiexec systems for job launch, via the ++mpiexec option to charmrun. Tested to work with no additional effort on Abe, and some small scripting on Ranger and Lonestar.
• Net-Linux: New charmrun option ++no-va-randomization to disable address space randomization (ASLR). This is most useful for running AMPI with isomalloc
• MPI: Default to using ampicxx instead of mpiCC
• MPI-SMP: The +p option now has the same semantics as in other smp builds
• Power 7: Support for VSX in SIMD abstraction API
• Blue Gene/L: Compilers and options have been updated to the latest ones
• Blue Gene/P
  • Added routines for measuring performance counters on BG/P.
  • Updated to support latest DCMF driver version. On ANL's Intrepid, you may need to set BGP_INSTALL=/bgsys/drivers/V1R4M1_460_2009-091110P/ppc in your environment. This is the default on ANL's Surveyor.
• Cray XT: cputopology information is now available on XT3/4/5
• Infiniband (ibverbs):
  • Bug fix: plug memory leaks that caused failures in long runs
  • Optimized to reduce startup delays
• LAPI: Support for SMP (experimental in 6.2.0, refined in 6.2.1)
• LAPI: Builds on Linux as well as AIX

Releases in the 6.2 series

• 6.2.0 (2010-04-06): First in release series
• 6.2.1 (2010-07-04): Bugfix and improvement patch release