forked from pmodels/casper
-
Notifications
You must be signed in to change notification settings - Fork 0
/
CHANGES
65 lines (51 loc) · 3.05 KB
/
CHANGES
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
======================================================================
Changes in 1.0b2
======================================================================
- Offloading-based asynchronous progress for MPI nonblocking two-sided
(point-to-point) communication. Enable this feature by setting
export CSP_ASYNC_MODE="pt2pt" or "pt2pt|rma" at executing time. This
environment variable is set to "rma" by default.
This feature relies on the openpa library.
(see http://trac.mpich.org/projects/openpa/)
We use the embedded openpa (commit 0475704d) by default. User can specify
an external openpa installation by setting --with-opa=[DIR] at configure
time.
- Performance profiling summary. This feature collects the amount of
RMA operations or two-sided messages that are handled by ghost processes
and the amount handled by user processes. This feature is disabled by
default because it inserts additional counter at performance critical
path. Setting --enable-profile at configure time to enable it.
- Hwloc based locality optimization. Casper automatically detects any
user core binding at MPI initialize time to ensure ghost processes are
evenly assigned onto each socket.
Example: if user sets core binding as 0,1,2,3 4,5,6,7, Casper will use
0 and 4 as the ghost process. In user-visible COMM_WORLD, rank 0-5 are
corresponding to 1-3,4-6, respectively.
This feature requires the hwloc library.
(see https://www.open-mpi.org/projects/hwloc/)
We embed hwloc version 1.11.8 in Casper by default. User can specify
an external hwloc installation by setting --with-hwloc=[DIR] at configure
time. To fully disable this feature, set --with-hwloc=no.
- Set CSP_NG=0 to fully disable Casper wrapping routines, and thus no
ghost process is isolated. Unlike setting CSP_NG=0, setting
CSP_ASYNC_CONFIG=off (on by default) can only disable RMA operation
redirection, and can be re-enabled for a window by setting
"async_config=on" info at window allocation time.
- BUG FIX: Support error handler for COMM_SELF, inter-communicator.
- BUG FIX: Fixed incompatible code with OpenMPI.
======================================================================
Changes in 1.0b1
======================================================================
- Provide asynchronous progress for RMA communication on window created
by MPI_Win_allocate.
- Runtime load balancing. This feature is disabled by default, enable it
by setting --enable-runtime-load at configure time.
- Multi-threading MPI supported. Internal critical sections are needed
to ensure thread-safety when using Casper with multi-threaded MPI
program (MPI_THREAD_MULTIPLE). This feature is disabled by default
for better performance, enable it by setting --enable-thread-safety
at configure time.
- IZEM thread library supported. IZEM provides light weight user-space
level thread lock. Use it for Casper internal critical sections instead
of traditional POSIX thread locks by setting --with-thread-package=izem
and --with-izem=[DIR] at configure time.