Skip to content
GitLab
Explore
Sign in
Primary navigation
Search or go to…
Project
S
swiftmpistepsim
Manage
Activity
Members
Labels
Plan
Issues
Issue boards
Milestones
Wiki
Code
Merge requests
Repository
Branches
Commits
Tags
Repository graph
Compare revisions
Snippets
Build
Pipelines
Jobs
Pipeline schedules
Artifacts
Deploy
Releases
Container registry
Model registry
Operate
Environments
Monitor
Incidents
Analyze
Value stream analytics
Contributor analytics
CI/CD analytics
Repository analytics
Model experiments
Help
Help
Support
GitLab documentation
Compare GitLab plans
GitLab community forum
Contribute to GitLab
Provide feedback
Keyboard shortcuts
?
Snippets
Groups
Projects
Show more breadcrumbs
SWIFT
swiftmpistepsim
Commits
ace196f8
Commit
ace196f8
authored
Aug 4, 2020
by
Peter W. Draper
Browse files
Options
Downloads
Patches
Plain Diff
Formatting
parent
1c2f92dd
Branches
Branches containing commit
No related tags found
1 merge request
!8
Draft: RDMA version with wrapped infinity calls
Changes
2
Show whitespace changes
Inline
Side-by-side
Showing
2 changed files
swiftmpirdmaonestepsim.c
+73
-79
73 additions, 79 deletions
swiftmpirdmaonestepsim.c
swiftmpirdmastepsim.c
+20
-20
20 additions, 20 deletions
swiftmpirdmastepsim.c
with
93 additions
and
99 deletions
swiftmpirdmaonestepsim.c
+
73
−
79
View file @
ace196f8
...
...
@@ -167,7 +167,9 @@ static BLOCKTYPE toblocks(BLOCKTYPE nr_bytes) {
*
* @result the number of bytes.
*/
static
BLOCKTYPE
tobytes
(
BLOCKTYPE
nr_blocks
)
{
return
(
nr_blocks
*
BYTESINBLOCK
);
}
static
BLOCKTYPE
tobytes
(
BLOCKTYPE
nr_blocks
)
{
return
(
nr_blocks
*
BYTESINBLOCK
);
}
/**
* @brief fill a data area with given value.
...
...
@@ -223,8 +225,7 @@ static void *send_thread(void *arg) {
/* Short circuit. */
if
(
nr_sends
[
rank
]
==
0
)
{
if
(
verbose
)
message
(
"took 0.0 %s."
,
clocks_getunit
());
if
(
verbose
)
message
(
"took 0.0 %s."
,
clocks_getunit
());
return
NULL
;
}
...
...
@@ -240,14 +241,13 @@ static void *send_thread(void *arg) {
server_ip
,
rank
,
BASE_PORT
+
myrank
);
auto
*
qp
=
qpFactory
->
connectToRemoteHost
(
server_ip
,
BASE_PORT
+
myrank
);
if
(
verbose
)
message
(
"%d connected to remote server %s %d on %d"
,
myrank
,
server_ip
,
rank
,
BASE_PORT
+
myrank
);
message
(
"%d connected to remote server %s %d on %d"
,
myrank
,
server_ip
,
rank
,
BASE_PORT
+
myrank
);
auto
*
remoteBufferToken
=
(
infinity
::
memory
::
RegionToken
*
)
qp
->
getUserData
();
/* Register some memory for use by RDMA, make it large enough for our
* biggest message. */
auto
*
sendBuffer
=
new
infinity
::
memory
::
Buffer
(
context
,
tobytes
(
max_size
));
auto
*
sendBuffer
=
new
infinity
::
memory
::
Buffer
(
context
,
tobytes
(
max_size
));
/* Queue of our sends. */
struct
mpiuse_log_entry
**
send_queue
=
sends_queue
[
rank
];
...
...
@@ -287,8 +287,8 @@ static void *send_thread(void *arg) {
for
(
size_t
j
=
0
;
j
<
counts
;
j
++
)
{
if
(
ranktag_lists
[
INDEX3
(
MAX_NR_RANKS
,
nr_ranks
,
log
->
rank
,
log
->
otherrank
,
j
)]
==
ranktag
)
{
log
->
offset
=
ranktag_offsets
[
INDEX3
(
MAX_NR_RANKS
,
nr_ranks
,
log
->
rank
,
log
->
otherrank
,
j
)];
log
->
offset
=
ranktag_offsets
[
INDEX3
(
MAX_NR_RANKS
,
nr_ranks
,
log
->
rank
,
log
->
otherrank
,
j
)];
found
=
1
;
break
;
}
...
...
@@ -297,12 +297,13 @@ static void *send_thread(void *arg) {
error
(
"Failed sending a message of size %zd to %d/%d "
"@ %zd
\n
, no offset found for ranktag %zd, counts = %zd"
,
datasize
,
log
->
otherrank
,
log
->
subtype
,
log
->
offset
,
ranktag
,
counts
);
datasize
,
log
->
otherrank
,
log
->
subtype
,
log
->
offset
,
ranktag
,
counts
);
}
if
(
verbose
)
message
(
"sending message subtype %d from %d to %d todo: %d/%d, offset %ld size %ld"
,
message
(
"sending message subtype %d from %d to %d todo: %d/%d, offset %ld "
"size %ld"
,
log
->
subtype
,
myrank
,
log
->
otherrank
,
nr_sends
[
rank
]
-
k
,
nr_sends
[
rank
],
log
->
offset
,
datasize
);
...
...
@@ -313,8 +314,7 @@ static void *send_thread(void *arg) {
remoteBufferToken
,
// destination
tobytes
(
log
->
offset
),
// remoteOffset
tobytes
(
datasize
),
// sizeInBytes
infinity
::
queues
::
OperationFlags
(),
&
requestToken
);
infinity
::
queues
::
OperationFlags
(),
&
requestToken
);
requestToken
.
waitUntilCompleted
();
requestToken
.
reset
();
...
...
@@ -325,20 +325,18 @@ static void *send_thread(void *arg) {
remoteBufferToken
,
// destination
tobytes
(
log
->
offset
),
// remoteOffset
BYTESINBLOCK
,
// sizeInBytes
infinity
::
queues
::
OperationFlags
(),
&
requestToken
);
infinity
::
queues
::
OperationFlags
(),
&
requestToken
);
requestToken
.
waitUntilCompleted
();
// Since we reuse the sendBuffer.
// requestToken.reset();
log
->
endtic
=
getticks
();
}
delete
sendBuffer
;
if
(
verbose
)
message
(
"sent %d to %d"
,
nr_sends
[
rank
],
rank
);
message
(
"took %.3f %s sending to rank %d"
,
clocks_from_ticks
(
getticks
()
-
starttics
),
clocks_getunit
(),
rank
);
if
(
verbose
)
message
(
"sent %d to %d"
,
nr_sends
[
rank
],
rank
);
message
(
"took %.3f %s sending to rank %d"
,
clocks_from_ticks
(
getticks
()
-
starttics
),
clocks_getunit
(),
rank
);
delete
sendBuffer
;
delete
qp
;
delete
qpFactory
;
delete
context
;
...
...
@@ -357,8 +355,7 @@ static void *recv_thread(void *arg) {
int
todo_recv
=
nr_recvs
[
rank
];
int
nr_recv
=
nr_recvs
[
rank
];
if
(
todo_recv
==
0
)
{
if
(
verbose
)
message
(
"took 0.0 %s."
,
clocks_getunit
());
if
(
verbose
)
message
(
"took 0.0 %s."
,
clocks_getunit
());
starting
[
rank
]
=
0
;
// Also fudge.
return
NULL
;
}
...
...
@@ -380,11 +377,10 @@ static void *recv_thread(void *arg) {
}
qpFactory
->
bindToPort
(
BASE_PORT
+
rank
);
if
(
verbose
)
message
(
"Blocking for first message on %d"
,
BASE_PORT
+
rank
);
if
(
verbose
)
message
(
"Blocking for first message on %d"
,
BASE_PORT
+
rank
);
starting
[
rank
]
=
0
;
// really need to do this in acceptIncomingConnection().
auto
qp
=
qpFactory
->
acceptIncomingConnection
(
bufferToken
,
sizeof
(
infinity
::
memory
::
RegionToken
));
auto
qp
=
qpFactory
->
acceptIncomingConnection
(
bufferToken
,
sizeof
(
infinity
::
memory
::
RegionToken
));
if
(
verbose
)
message
(
"Accepting incoming connections on %d"
,
BASE_PORT
+
rank
);
...
...
@@ -421,7 +417,8 @@ static void *recv_thread(void *arg) {
// int tag = dataptr[3];
if
(
verbose
)
message
(
"receive message subtype %d from %d to %d todo: %d/%d,"
message
(
"receive message subtype %d from %d to %d todo: %d/%d,"
" offset %ld size %ld"
,
log
->
subtype
,
myrank
,
log
->
otherrank
,
todo_recv
,
nr_recv
,
log
->
offset
,
toblocks
(
log
->
size
)
+
HEADER_SIZE
);
...
...
@@ -442,13 +439,14 @@ static void *recv_thread(void *arg) {
}
}
message
(
"took %.3f %s listening for rank %d"
,
clocks_from_ticks
(
getticks
()
-
starttics
),
clocks_getunit
(),
rank
);
message
(
"took %.3f %s listening for rank %d"
,
clocks_from_ticks
(
getticks
()
-
starttics
),
clocks_getunit
(),
rank
);
delete
qp
;
// XXX Let this leak, tearing it down can be premature for the remote.
// delete qp;
qp
->
getLocalDeviceId
();
// Stop compiler complaining.
delete
qpFactory
;
/* Thread exits. */
return
NULL
;
}
...
...
@@ -474,11 +472,11 @@ static size_t pick_logs() {
size_t
maxsize
=
0
;
/* Queues of send and receive logs. */
struct
mpiuse_log_entry
**
send_queue
=
(
struct
mpiuse_log_entry
**
)
calloc
(
nlogs
,
sizeof
(
struct
mpiuse_log_entry
*
));
struct
mpiuse_log_entry
**
send_queue
=
(
struct
mpiuse_log_entry
**
)
calloc
(
nlogs
,
sizeof
(
struct
mpiuse_log_entry
*
));
int
nr_send
=
0
;
struct
mpiuse_log_entry
**
recv_queue
=
(
struct
mpiuse_log_entry
**
)
calloc
(
nlogs
,
sizeof
(
struct
mpiuse_log_entry
*
));
struct
mpiuse_log_entry
**
recv_queue
=
(
struct
mpiuse_log_entry
**
)
calloc
(
nlogs
,
sizeof
(
struct
mpiuse_log_entry
*
));
int
nr_recv
=
0
;
for
(
size_t
k
=
0
;
k
<
nlogs
;
k
++
)
{
...
...
@@ -533,8 +531,8 @@ static size_t pick_logs() {
/* Recvs */
if
(
nr_recvs
[
k
]
>
0
)
{
recvs_queue
[
k
]
=
(
struct
mpiuse_log_entry
**
)
calloc
(
nr_recvs
[
k
],
sizeof
(
struct
mpiuse_log_entry
*
));
recvs_queue
[
k
]
=
(
struct
mpiuse_log_entry
**
)
calloc
(
nr_recvs
[
k
],
sizeof
(
struct
mpiuse_log_entry
*
));
int
i
=
0
;
for
(
int
j
=
0
;
j
<
nr_recv
;
j
++
)
{
struct
mpiuse_log_entry
*
log
=
recv_queue
[
j
];
...
...
@@ -543,15 +541,14 @@ static size_t pick_logs() {
i
++
;
}
}
}
else
{
}
else
{
recvs_queue
[
k
]
=
NULL
;
}
/* Sends. */
if
(
nr_sends
[
k
]
>
0
)
{
sends_queue
[
k
]
=
(
struct
mpiuse_log_entry
**
)
calloc
(
nr_sends
[
k
],
sizeof
(
struct
mpiuse_log_entry
*
));
sends_queue
[
k
]
=
(
struct
mpiuse_log_entry
**
)
calloc
(
nr_sends
[
k
],
sizeof
(
struct
mpiuse_log_entry
*
));
int
i
=
0
;
for
(
int
j
=
0
;
j
<
nr_send
;
j
++
)
{
struct
mpiuse_log_entry
*
log
=
send_queue
[
j
];
...
...
@@ -560,17 +557,16 @@ static size_t pick_logs() {
i
++
;
}
}
}
else
{
}
else
{
sends_queue
[
k
]
=
NULL
;
}
}
/* Offsets and ranktags. */
ranktag_offsets
=
(
size_t
*
)
calloc
(
MAX_NR_RANKS
*
nr_ranks
*
max_logs
,
sizeof
(
size_t
));
ranktag_lists
=
(
size_t
*
)
calloc
(
MAX_NR_RANKS
*
nr_ranks
*
max_logs
,
sizeof
(
size_t
));
ranktag_offsets
=
(
size_t
*
)
calloc
(
MAX_NR_RANKS
*
nr_ranks
*
max_logs
,
sizeof
(
size_t
));
ranktag_lists
=
(
size_t
*
)
calloc
(
MAX_NR_RANKS
*
nr_ranks
*
max_logs
,
sizeof
(
size_t
));
// ranktag_counts = (size_t *)calloc(MAX_NR_RANKS * nr_ranks, sizeof(size_t));
/* Setup the ranktag offsets for our receive windows. Also define the sizes
...
...
@@ -692,9 +688,8 @@ int main(int argc, char *argv[]) {
MPI_COMM_WORLD
);
// MPI_Allreduce(MPI_IN_PLACE, ranktag_counts, MAX_NR_RANKS * nr_ranks,
// MPI_AINT, MPI_SUM, MPI_COMM_WORLD);
MPI_Allreduce
(
MPI_IN_PLACE
,
ranktag_lists
,
MAX_NR_RANKS
*
nr_ranks
*
max_logs
,
MPI_AINT
,
MPI_SUM
,
MPI_COMM_WORLD
);
MPI_Allreduce
(
MPI_IN_PLACE
,
ranktag_lists
,
MAX_NR_RANKS
*
nr_ranks
*
max_logs
,
MPI_AINT
,
MPI_SUM
,
MPI_COMM_WORLD
);
/* Now for the RDMA setup. We need the IP addresses of all the ranks. */
...
...
@@ -762,8 +757,7 @@ int main(int argc, char *argv[]) {
/* Reset time as previous can be thought of as setup costs? */
MPI_Barrier
(
MPI_COMM_WORLD
);
// Vital...
if
(
myrank
==
0
)
message
(
"All synchronized, restarting time."
);
if
(
myrank
==
0
)
message
(
"All synchronized, restarting time."
);
clocks_set_cpufreq
(
0
);
/* Now we have a thread per rank to send the messages. */
...
...
...
...
This diff is collapsed.
Click to expand it.
swiftmpirdmastepsim.c
+
20
−
20
View file @
ace196f8
...
...
@@ -127,7 +127,9 @@ static int toblocks(BLOCKTYPE nr_bytes) {
*
* @result the number of bytes.
*/
static
BLOCKTYPE
tobytes
(
BLOCKTYPE
nr_blocks
)
{
return
(
nr_blocks
*
BYTESINBLOCK
);
}
static
BLOCKTYPE
tobytes
(
BLOCKTYPE
nr_blocks
)
{
return
(
nr_blocks
*
BYTESINBLOCK
);
}
/**
* @brief fill a data area with given value.
...
...
@@ -191,7 +193,8 @@ static void *send_thread(void *arg) {
// message("%d waiting for connection to remote server %s %d on %d", myrank,
// server_ip, rank, BASE_PORT + myrank);
auto
*
qp
=
qpFactory
->
connectToRemoteHost
(
server_ip
,
BASE_PORT
+
myrank
);
//message("%d connected to remote server %s %d on %d", myrank, server_ip, rank,
// message("%d connected to remote server %s %d on %d", myrank, server_ip,
// rank,
// BASE_PORT + myrank);
for
(
int
k
=
0
;
k
<
nr_send
;
k
++
)
{
...
...
@@ -218,8 +221,8 @@ static void *send_thread(void *arg) {
dataptr
[
3
]
=
log
->
tag
;
/* Need to assign to a buffer to register memory. */
auto
*
sendBuffer
=
new
infinity
::
memory
::
Buffer
(
context
,
dataptr
,
tobytes
(
datasize
));
auto
*
sendBuffer
=
new
infinity
::
memory
::
Buffer
(
context
,
dataptr
,
tobytes
(
datasize
));
// And send
infinity
::
requests
::
RequestToken
requestToken
(
context
);
...
...
@@ -263,8 +266,7 @@ static void *recv_thread(void *arg) {
message
(
"%d binding to %d on port %d"
,
myrank
,
rank
,
BASE_PORT
+
rank
);
qpFactory
->
bindToPort
(
BASE_PORT
+
rank
);
if
(
verbose
)
message
(
"Blocking for first message on %d"
,
BASE_PORT
+
rank
);
if
(
verbose
)
message
(
"Blocking for first message on %d"
,
BASE_PORT
+
rank
);
starting
[
rank
]
=
0
;
// really need to do this in acceptIncomingConnection().
auto
qp
=
qpFactory
->
acceptIncomingConnection
();
if
(
verbose
)
...
...
@@ -365,11 +367,11 @@ static size_t pick_logs() {
size_t
maxsize
=
0
;
/* Queues of send and receive logs. */
send_queue
=
(
struct
mpiuse_log_entry
**
)
calloc
(
nlogs
,
sizeof
(
struct
mpiuse_log_entry
*
));
send_queue
=
(
struct
mpiuse_log_entry
**
)
calloc
(
nlogs
,
sizeof
(
struct
mpiuse_log_entry
*
));
nr_send
=
0
;
struct
mpiuse_log_entry
**
recv_queue
=
(
struct
mpiuse_log_entry
**
)
calloc
(
nlogs
,
sizeof
(
struct
mpiuse_log_entry
*
));
struct
mpiuse_log_entry
**
recv_queue
=
(
struct
mpiuse_log_entry
**
)
calloc
(
nlogs
,
sizeof
(
struct
mpiuse_log_entry
*
));
int
nr_recv
=
0
;
for
(
size_t
k
=
0
;
k
<
nlogs
;
k
++
)
{
...
...
@@ -413,8 +415,8 @@ static size_t pick_logs() {
}
for
(
int
k
=
0
;
k
<
nr_ranks
;
k
++
)
{
if
(
nr_recvs
[
k
]
>
0
)
{
recvs_queue
[
k
]
=
(
struct
mpiuse_log_entry
**
)
calloc
(
nr_recvs
[
k
],
sizeof
(
struct
mpiuse_log_entry
*
));
recvs_queue
[
k
]
=
(
struct
mpiuse_log_entry
**
)
calloc
(
nr_recvs
[
k
],
sizeof
(
struct
mpiuse_log_entry
*
));
int
i
=
0
;
for
(
int
j
=
0
;
j
<
nr_recv
;
j
++
)
{
struct
mpiuse_log_entry
*
log
=
recv_queue
[
j
];
...
...
@@ -423,14 +425,12 @@ static size_t pick_logs() {
i
++
;
}
}
}
else
{
}
else
{
recvs_queue
[
k
]
=
NULL
;
}
}
free
(
recv_queue
);
if
(
verbose
)
{
message
(
"maxsize = %zd, nr_send = %d, nr_recv = %d"
,
maxsize
,
nr_send
,
nr_recv
);
...
...
...
...
This diff is collapsed.
Click to expand it.
Preview
0%
Loading
Try again
or
attach a new file
.
Cancel
You are about to add
0
people
to the discussion. Proceed with caution.
Finish editing this message first!
Save comment
Cancel
Please
sign in
to comment