[OMPI users] Latencies of atomic operations on high-performance networks

[Joseph Schuchart]

All,

I am currently experimenting with MPI atomic operations and wanted to
share some interesting results I am observing. The numbers below are
measurements from both an IB-based cluster and our Cray XC40. The
benchmarks look like the following snippet:

```
if (rank == 1) {
uint64_t res, val;
for (size_t i = 0; i < NUM_REPS; ++i) {
MPI_Fetch_and_op(&val, &res, MPI_UINT32_T, 0, 0, MPI_SUM, win);
MPI_Win_flush(target, win);
}
}
MPI_Barrier(MPI_COMM_WORLD);
```

Only rank 1 performs atomic operations, rank 0 waits in a barrier (I
have tried to confirm that the operations are done in hardware by
letting rank 0 sleep for a while and ensuring that communication
progresses). Of particular interest for my use-case is fetch_op but I am
including other operations here nevertheless:

* Linux Cluster, IB QDR *
average of 100000 iterations

Exclusive lock, MPI_UINT32_T:
fetch_op: 4.323384us
compare_exchange: 2.035905us
accumulate: 4.326358us
get_accumulate: 4.334831us

Exclusive lock, MPI_UINT64_T:
fetch_op: 2.438080us
compare_exchange: 2.398836us
accumulate: 2.435378us
get_accumulate: 2.448347us

Shared lock, MPI_UINT32_T:
fetch_op: 6.819977us
compare_exchange: 4.551417us
accumulate: 6.807766us
get_accumulate: 6.817602us

Shared lock, MPI_UINT64_T:
fetch_op: 4.954860us
compare_exchange: 2.399373us
accumulate: 4.965702us
get_accumulate: 4.977876us

There are two interesting observations:
a) operations on 64bit operands generally seem to have lower latencies
than operations on 32bit
b) Using an exclusive lock leads to lower latencies

Overall, there is a factor of almost 3 between SharedLock+uint32_t and
ExclusiveLock+uint64_t for fetch_and_op, accumulate, and get_accumulate
(compare_exchange seems to be somewhat of an outlier).

* Cray XC40, Aries *
average of 100000 iterations

Exclusive lock, MPI_UINT32_T:
fetch_op: 2.011794us
compare_exchange: 1.740825us
accumulate: 1.795500us
get_accumulate: 1.985409us

Exclusive lock, MPI_UINT64_T:
fetch_op: 2.017172us
compare_exchange: 1.846202us
accumulate: 1.812578us
get_accumulate: 2.005541us

Shared lock, MPI_UINT32_T:
fetch_op: 5.380455us
compare_exchange: 5.164458us
accumulate: 5.230184us
get_accumulate: 5.399722us

Shared lock, MPI_UINT64_T:
fetch_op: 5.415230us
compare_exchange: 1.855840us
accumulate: 5.212632us
get_accumulate: 5.396110us


The difference between exclusive and shared lock is about the same as
with IB and the latencies for 32bit vs 64bit are roughly the same
(except for compare_exchange, it seems).

So my question is: is this to be expected? Is the higher latency when
using a shared lock caused by an internal lock being acquired because
the hardware operations are not actually atomic?

I'd be grateful for any insight on this.

Cheers,
Joseph

[Nathan Hjelm via users]

All of this is completely expected. Due to the requirements of the standard it is difficult to make use of network atomics even for MPI_Compare_and_swap (MPI_Accumulate and MPI_Get_accumulate spoil the party). If you want MPI_Fetch_and_op to be fast set this MCA parameter:


osc_rdma_acc_single_intrinsic=true


Shared lock is slower than an exclusive lock because there is an extra lock step as part of the accumulate (it isn't needed if there is an exclusive lock). When setting the above parameter you are telling the implementation that you will only be using a single count and we can optimize that with the hardware. The RMA working group is working on an info key that will essentially do the same thing.


Note the above parameter won't help you with IB if you are using UCX unless you set this (master only right now):



btl_uct_transports=dc_mlx5

btl=self,vader,uct

osc=^ucx




Though there may be a way to get osc/ucx to enable the same sort of optimization. I don't know.



-Nathan



On Nov 06, 2018, at 09:38 AM, Joseph Schuchart <***@hlrs.de> wrote:


All,

I am currently experimenting with MPI atomic operations and wanted to
share some interesting results I am observing. The numbers below are
measurements from both an IB-based cluster and our Cray XC40. The
benchmarks look like the following snippet:

```
if (rank == 1) {
uint64_t res, val;
for (size_t i = 0; i < NUM_REPS; ++i) {
MPI_Fetch_and_op(&val, &res, MPI_UINT32_T, 0, 0, MPI_SUM, win);
MPI_Win_flush(target, win);
}
}
MPI_Barrier(MPI_COMM_WORLD);
```

Only rank 1 performs atomic operations, rank 0 waits in a barrier (I
have tried to confirm that the operations are done in hardware by
letting rank 0 sleep for a while and ensuring that communication
progresses). Of particular interest for my use-case is fetch_op but I am
including other operations here nevertheless:

* Linux Cluster, IB QDR *
average of 100000 iterations

Exclusive lock, MPI_UINT32_T:
fetch_op: 4.323384us
compare_exchange: 2.035905us
accumulate: 4.326358us
get_accumulate: 4.334831us

Exclusive lock, MPI_UINT64_T:
fetch_op: 2.438080us
compare_exchange: 2.398836us
accumulate: 2.435378us
get_accumulate: 2.448347us

Shared lock, MPI_UINT32_T:
fetch_op: 6.819977us
compare_exchange: 4.551417us
accumulate: 6.807766us
get_accumulate: 6.817602us

Shared lock, MPI_UINT64_T:
fetch_op: 4.954860us
compare_exchange: 2.399373us
accumulate: 4.965702us
get_accumulate: 4.977876us

There are two interesting observations:
a) operations on 64bit operands generally seem to have lower latencies
than operations on 32bit
b) Using an exclusive lock leads to lower latencies

Overall, there is a factor of almost 3 between SharedLock+uint32_t and
ExclusiveLock+uint64_t for fetch_and_op, accumulate, and get_accumulate
(compare_exchange seems to be somewhat of an outlier).

* Cray XC40, Aries *
average of 100000 iterations

Exclusive lock, MPI_UINT32_T:
fetch_op: 2.011794us
compare_exchange: 1.740825us
accumulate: 1.795500us
get_accumulate: 1.985409us

Exclusive lock, MPI_UINT64_T:
fetch_op: 2.017172us
compare_exchange: 1.846202us
accumulate: 1.812578us
get_accumulate: 2.005541us

Shared lock, MPI_UINT32_T:
fetch_op: 5.380455us
compare_exchange: 5.164458us
accumulate: 5.230184us
get_accumulate: 5.399722us

Shared lock, MPI_UINT64_T:
fetch_op: 5.415230us
compare_exchange: 1.855840us
accumulate: 5.212632us
get_accumulate: 5.396110us


The difference between exclusive and shared lock is about the same as
with IB and the latencies for 32bit vs 64bit are roughly the same
(except for compare_exchange, it seems).

So my question is: is this to be expected? Is the higher latency when
using a shared lock caused by an internal lock being acquired because
the hardware operations are not actually atomic?

I'd be grateful for any insight on this.

Cheers,
Joseph

--
Dipl.-Inf. Joseph Schuchart
High Performance Computing Center Stuttgart (HLRS)
Nobelstr. 19
D-70569 Stuttgart

Tel.: [phone]
Fax: [phone]
E-Mail: ***@hlrs.de

[Joseph Schuchart]

Thanks a lot for the quick reply, setting
osc_rdma_acc_single_intrinsic=true does the trick for both shared and
exclusive locks and brings it down to <2us per operation. I hope that
the info key will make it into the next version of the standard, I
certainly have use for it :)

Cheers,
Joseph

Post by Nathan Hjelm via users
All of this is completely expected. Due to the requirements of the
standard it is difficult to make use of network atomics even for
MPI_Compare_and_swap (MPI_Accumulate and MPI_Get_accumulate spoil the
osc_rdma_acc_single_intrinsic=true
Shared lock is slower than an exclusive lock because there is an extra
lock step as part of the accumulate (it isn't needed if there is an
exclusive lock). When setting the above parameter you are telling the
implementation that you will only be using a single count and we can
optimize that with the hardware. The RMA working group is working on an
info key that will essentially do the same thing.
Note the above parameter won't help you with IB if you are using UCX
btl_uct_transports=dc_mlx5
btl=self,vader,uct
osc=^ucx
Though there may be a way to get osc/ucx to enable the same sort of
optimization. I don't know.
-Nathan

Post by Joseph Schuchart
All,
I am currently experimenting with MPI atomic operations and wanted to
share some interesting results I am observing. The numbers below are
measurements from both an IB-based cluster and our Cray XC40. The
```
if (rank == 1) {
uint64_t res, val;
for (size_t i = 0; i < NUM_REPS; ++i) {
MPI_Fetch_and_op(&val, &res, MPI_UINT32_T, 0, 0, MPI_SUM, win);
MPI_Win_flush(target, win);
}
}
MPI_Barrier(MPI_COMM_WORLD);
```
Only rank 1 performs atomic operations, rank 0 waits in a barrier (I
have tried to confirm that the operations are done in hardware by
letting rank 0 sleep for a while and ensuring that communication
progresses). Of particular interest for my use-case is fetch_op but I am
* Linux Cluster, IB QDR *
average of 100000 iterations
fetch_op: 4.323384us
compare_exchange: 2.035905us
accumulate: 4.326358us
get_accumulate: 4.334831us
fetch_op: 2.438080us
compare_exchange: 2.398836us
accumulate: 2.435378us
get_accumulate: 2.448347us
fetch_op: 6.819977us
compare_exchange: 4.551417us
accumulate: 6.807766us
get_accumulate: 6.817602us
fetch_op: 4.954860us
compare_exchange: 2.399373us
accumulate: 4.965702us
get_accumulate: 4.977876us
a) operations on 64bit operands generally seem to have lower latencies
than operations on 32bit
b) Using an exclusive lock leads to lower latencies
Overall, there is a factor of almost 3 between SharedLock+uint32_t and
ExclusiveLock+uint64_t for fetch_and_op, accumulate, and get_accumulate
(compare_exchange seems to be somewhat of an outlier).
* Cray XC40, Aries *
average of 100000 iterations
fetch_op: 2.011794us
compare_exchange: 1.740825us
accumulate: 1.795500us
get_accumulate: 1.985409us
fetch_op: 2.017172us
compare_exchange: 1.846202us
accumulate: 1.812578us
get_accumulate: 2.005541us
fetch_op: 5.380455us
compare_exchange: 5.164458us
accumulate: 5.230184us
get_accumulate: 5.399722us
fetch_op: 5.415230us
compare_exchange: 1.855840us
accumulate: 5.212632us
get_accumulate: 5.396110us
The difference between exclusive and shared lock is about the same as
with IB and the latencies for 32bit vs 64bit are roughly the same
(except for compare_exchange, it seems).
So my question is: is this to be expected? Is the higher latency when
using a shared lock caused by an internal lock being acquired because
the hardware operations are not actually atomic?
I'd be grateful for any insight on this.
Cheers,
Joseph
--
Dipl.-Inf. Joseph Schuchart
High Performance Computing Center Stuttgart (HLRS)
Nobelstr. 19
D-70569 Stuttgart
Tel.: [phone]
Fax: [phone]
_______________________________________________
users mailing list
https://lists.open-mpi.org/mailman/listinfo/users

_______________________________________________
users mailing list
https://lists.open-mpi.org/mailman/listinfo/users

[Joseph Schuchart]

While using the mca parameter in a real application I noticed a strange
effect, which took me a while to figure out: It appears that on the
Aries network the accumulate operations are not atomic anymore. I am
attaching a test program that shows the problem: all but one processes
continuously increment a counter while rank 0 is continuously
subtracting a large value and adding it again, eventually checking for
the correct number of increments. Without the mca parameter the test at
the end succeeds as all increments are accounted for:

```
$ mpirun -n 16 -N 1 ./mpi_fetch_op_local_remote
result:15000
```

When setting the mca parameter the test fails with garbage in the result:

```
$ mpirun --mca osc_rdma_acc_single_intrinsic true -n 16 -N 1
./mpi_fetch_op_local_remote
result:25769849013
mpi_fetch_op_local_remote: mpi_fetch_op_local_remote.c:97: main:
Assertion `sum == 1000*(comm_size-1)' failed.
```

All processes perform only MPI_Fetch_and_op in combination with MPI_SUM
so I assume that the test in combination with the mca flag is correct. I
cannot reproduce this issue on our IB cluster.

Is that an issue in Open MPI or is there some problem in the test case
that I am missing?

Thanks in advance,
Joseph

Post by Joseph Schuchart
Thanks a lot for the quick reply, setting
osc_rdma_acc_single_intrinsic=true does the trick for both shared and
exclusive locks and brings it down to <2us per operation. I hope that
the info key will make it into the next version of the standard, I
certainly have use for it :)
Cheers,
Joseph

Post by Nathan Hjelm via users
All of this is completely expected. Due to the requirements of the
standard it is difficult to make use of network atomics even for
MPI_Compare_and_swap (MPI_Accumulate and MPI_Get_accumulate spoil the
osc_rdma_acc_single_intrinsic=true
Shared lock is slower than an exclusive lock because there is an extra
lock step as part of the accumulate (it isn't needed if there is an
exclusive lock). When setting the above parameter you are telling the
implementation that you will only be using a single count and we can
optimize that with the hardware. The RMA working group is working on
an info key that will essentially do the same thing.
Note the above parameter won't help you with IB if you are using UCX
btl_uct_transports=dc_mlx5
btl=self,vader,uct
osc=^ucx
Though there may be a way to get osc/ucx to enable the same sort of
optimization. I don't know.
-Nathan

Post by Joseph Schuchart
All,
I am currently experimenting with MPI atomic operations and wanted to
share some interesting results I am observing. The numbers below are
measurements from both an IB-based cluster and our Cray XC40. The
```
if (rank == 1) {
uint64_t res, val;
for (size_t i = 0; i < NUM_REPS; ++i) {
MPI_Fetch_and_op(&val, &res, MPI_UINT32_T, 0, 0, MPI_SUM, win);
MPI_Win_flush(target, win);
}
}
MPI_Barrier(MPI_COMM_WORLD);
```
Only rank 1 performs atomic operations, rank 0 waits in a barrier (I
have tried to confirm that the operations are done in hardware by
letting rank 0 sleep for a while and ensuring that communication
progresses). Of particular interest for my use-case is fetch_op but I am
* Linux Cluster, IB QDR *
average of 100000 iterations
fetch_op: 4.323384us
compare_exchange: 2.035905us
accumulate: 4.326358us
get_accumulate: 4.334831us
fetch_op: 2.438080us
compare_exchange: 2.398836us
accumulate: 2.435378us
get_accumulate: 2.448347us
fetch_op: 6.819977us
compare_exchange: 4.551417us
accumulate: 6.807766us
get_accumulate: 6.817602us
fetch_op: 4.954860us
compare_exchange: 2.399373us
accumulate: 4.965702us
get_accumulate: 4.977876us
a) operations on 64bit operands generally seem to have lower latencies
than operations on 32bit
b) Using an exclusive lock leads to lower latencies
Overall, there is a factor of almost 3 between SharedLock+uint32_t and
ExclusiveLock+uint64_t for fetch_and_op, accumulate, and get_accumulate
(compare_exchange seems to be somewhat of an outlier).
* Cray XC40, Aries *
average of 100000 iterations
fetch_op: 2.011794us
compare_exchange: 1.740825us
accumulate: 1.795500us
get_accumulate: 1.985409us
fetch_op: 2.017172us
compare_exchange: 1.846202us
accumulate: 1.812578us
get_accumulate: 2.005541us
fetch_op: 5.380455us
compare_exchange: 5.164458us
accumulate: 5.230184us
get_accumulate: 5.399722us
fetch_op: 5.415230us
compare_exchange: 1.855840us
accumulate: 5.212632us
get_accumulate: 5.396110us
The difference between exclusive and shared lock is about the same as
with IB and the latencies for 32bit vs 64bit are roughly the same
(except for compare_exchange, it seems).
So my question is: is this to be expected? Is the higher latency when
using a shared lock caused by an internal lock being acquired because
the hardware operations are not actually atomic?
I'd be grateful for any insight on this.
Cheers,
Joseph
--
Dipl.-Inf. Joseph Schuchart
High Performance Computing Center Stuttgart (HLRS)
Nobelstr. 19
D-70569 Stuttgart
Tel.: [phone]
Fax: [phone]
_______________________________________________
users mailing list
https://lists.open-mpi.org/mailman/listinfo/users

_______________________________________________
users mailing list
https://lists.open-mpi.org/mailman/listinfo/users

_______________________________________________
users mailing list
https://lists.open-mpi.org/mailman/listinfo/users

[Nathan Hjelm via users]

Quick scan of the program and it looks ok to me. I will dig deeper and see if I can determine the underlying cause.


What Open MPI version are you using?


-Nathan

On Nov 08, 2018, at 11:10 AM, Joseph Schuchart <***@hlrs.de> wrote:


While using the mca parameter in a real application I noticed a strange
effect, which took me a while to figure out: It appears that on the
Aries network the accumulate operations are not atomic anymore. I am
attaching a test program that shows the problem: all but one processes
continuously increment a counter while rank 0 is continuously
subtracting a large value and adding it again, eventually checking for
the correct number of increments. Without the mca parameter the test at
the end succeeds as all increments are accounted for:

```
$ mpirun -n 16 -N 1 ./mpi_fetch_op_local_remote
result:15000
```

When setting the mca parameter the test fails with garbage in the result:

```
$ mpirun --mca osc_rdma_acc_single_intrinsic true -n 16 -N 1
./mpi_fetch_op_local_remote
result:25769849013
mpi_fetch_op_local_remote: mpi_fetch_op_local_remote.c:97: main:
Assertion `sum == 1000*(comm_size-1)' failed.
```

All processes perform only MPI_Fetch_and_op in combination with MPI_SUM
so I assume that the test in combination with the mca flag is correct. I
cannot reproduce this issue on our IB cluster.

Is that an issue in Open MPI or is there some problem in the test case
that I am missing?

Thanks in advance,
Joseph


On 11/6/18 1:15 PM, Joseph Schuchart wrote:

Thanks a lot for the quick reply, setting
osc_rdma_acc_single_intrinsic=true does the trick for both shared and
exclusive locks and brings it down to <2us per operation. I hope that
the info key will make it into the next version of the standard, I
certainly have use for it :)


Cheers,
Joseph


On 11/6/18 12:13 PM, Nathan Hjelm via users wrote:


All of this is completely expected. Due to the requirements of the
standard it is difficult to make use of network atomics even for
MPI_Compare_and_swap (MPI_Accumulate and MPI_Get_accumulate spoil the
party). If you want MPI_Fetch_and_op to be fast set this MCA parameter:


osc_rdma_acc_single_intrinsic=true


Shared lock is slower than an exclusive lock because there is an extra
lock step as part of the accumulate (it isn't needed if there is an
exclusive lock). When setting the above parameter you are telling the
implementation that you will only be using a single count and we can
optimize that with the hardware. The RMA working group is working on
an info key that will essentially do the same thing.


Note the above parameter won't help you with IB if you are using UCX
unless you set this (master only right now):


btl_uct_transports=dc_mlx5


btl=self,vader,uct


osc=^ucx




Though there may be a way to get osc/ucx to enable the same sort of
optimization. I don't know.




-Nathan




On Nov 06, 2018, at 09:38 AM, Joseph Schuchart <***@hlrs.de> wrote:


All,


I am currently experimenting with MPI atomic operations and wanted to
share some interesting results I am observing. The numbers below are
measurements from both an IB-based cluster and our Cray XC40. The
benchmarks look like the following snippet:


```
if (rank == 1) {
uint64_t res, val;
for (size_t i = 0; i < NUM_REPS; ++i) {
MPI_Fetch_and_op(&val, &res, MPI_UINT32_T, 0, 0, MPI_SUM, win);
MPI_Win_flush(target, win);
}
}
MPI_Barrier(MPI_COMM_WORLD);
```


Only rank 1 performs atomic operations, rank 0 waits in a barrier (I
have tried to confirm that the operations are done in hardware by
letting rank 0 sleep for a while and ensuring that communication
progresses). Of particular interest for my use-case is fetch_op but I am
including other operations here nevertheless:


* Linux Cluster, IB QDR *
average of 100000 iterations


Exclusive lock, MPI_UINT32_T:
fetch_op: 4.323384us
compare_exchange: 2.035905us
accumulate: 4.326358us
get_accumulate: 4.334831us


Exclusive lock, MPI_UINT64_T:
fetch_op: 2.438080us
compare_exchange: 2.398836us
accumulate: 2.435378us
get_accumulate: 2.448347us


Shared lock, MPI_UINT32_T:
fetch_op: 6.819977us
compare_exchange: 4.551417us
accumulate: 6.807766us
get_accumulate: 6.817602us


Shared lock, MPI_UINT64_T:
fetch_op: 4.954860us
compare_exchange: 2.399373us
accumulate: 4.965702us
get_accumulate: 4.977876us


There are two interesting observations:
a) operations on 64bit operands generally seem to have lower latencies
than operations on 32bit
b) Using an exclusive lock leads to lower latencies


Overall, there is a factor of almost 3 between SharedLock+uint32_t and
ExclusiveLock+uint64_t for fetch_and_op, accumulate, and get_accumulate
(compare_exchange seems to be somewhat of an outlier).


* Cray XC40, Aries *
average of 100000 iterations


Exclusive lock, MPI_UINT32_T:
fetch_op: 2.011794us
compare_exchange: 1.740825us
accumulate: 1.795500us
get_accumulate: 1.985409us


Exclusive lock, MPI_UINT64_T:
fetch_op: 2.017172us
compare_exchange: 1.846202us
accumulate: 1.812578us
get_accumulate: 2.005541us


Shared lock, MPI_UINT32_T:
fetch_op: 5.380455us
compare_exchange: 5.164458us
accumulate: 5.230184us
get_accumulate: 5.399722us


Shared lock, MPI_UINT64_T:
fetch_op: 5.415230us
compare_exchange: 1.855840us
accumulate: 5.212632us
get_accumulate: 5.396110us




The difference between exclusive and shared lock is about the same as
with IB and the latencies for 32bit vs 64bit are roughly the same
(except for compare_exchange, it seems).


So my question is: is this to be expected? Is the higher latency when
using a shared lock caused by an internal lock being acquired because
the hardware operations are not actually atomic?


I'd be grateful for any insight on this.


Cheers,
Joseph


--
Dipl.-Inf. Joseph Schuchart
High Performance Computing Center Stuttgart (HLRS)
Nobelstr. 19
D-70569 Stuttgart


Tel.: [phone]
Fax: [phone]
E-Mail: ***@hlrs.de <mailto:***@hlrs.de>
_______________________________________________
users mailing list
***@lists.open-mpi.org <mailto:***@lists.open-mpi.org>
https://lists.open-mpi.org/mailman/listinfo/users


_______________________________________________
users mailing list
***@lists.open-mpi.org
https://lists.open-mpi.org/mailman/listinfo/users

[Joseph Schuchart]

Sorry for the delay, I wanted to make sure that I test the same version
on both Aries and IB: git master bbe5da4. I realized that I had
previously tested with 3.1.3 on the IB cluster, which ran fine. If I use
the same version I run into the same problem on both systems (with --mca
btl_openib_allow_ib true --mca osc_rdma_acc_single_intrinsic true). I
have not tried using UCX for this.

Joseph

Post by Nathan Hjelm via users
Quick scan of the program and it looks ok to me. I will dig deeper and
see if I can determine the underlying cause.
What Open MPI version are you using?
-Nathan

Post by Joseph Schuchart
While using the mca parameter in a real application I noticed a strange
effect, which took me a while to figure out: It appears that on the
Aries network the accumulate operations are not atomic anymore. I am
attaching a test program that shows the problem: all but one processes
continuously increment a counter while rank 0 is continuously
subtracting a large value and adding it again, eventually checking for
the correct number of increments. Without the mca parameter the test at
```
$ mpirun -n 16 -N 1 ./mpi_fetch_op_local_remote
result:15000
```
```
$ mpirun --mca osc_rdma_acc_single_intrinsic true -n 16 -N 1
./mpi_fetch_op_local_remote
result:25769849013
Assertion `sum == 1000*(comm_size-1)' failed.
```
All processes perform only MPI_Fetch_and_op in combination with MPI_SUM
so I assume that the test in combination with the mca flag is correct. I
cannot reproduce this issue on our IB cluster.
Is that an issue in Open MPI or is there some problem in the test case
that I am missing?
Thanks in advance,
Joseph

Post by Joseph Schuchart
Thanks a lot for the quick reply, setting
osc_rdma_acc_single_intrinsic=true does the trick for both shared and
exclusive locks and brings it down to <2us per operation. I hope that
the info key will make it into the next version of the standard, I
certainly have use for it :)
Cheers,
Joseph

Post by Nathan Hjelm via users
All of this is completely expected. Due to the requirements of the
standard it is difficult to make use of network atomics even for
MPI_Compare_and_swap (MPI_Accumulate and MPI_Get_accumulate spoil the
osc_rdma_acc_single_intrinsic=true
Shared lock is slower than an exclusive lock because there is an extra
lock step as part of the accumulate (it isn't needed if there is an
exclusive lock). When setting the above parameter you are telling the
implementation that you will only be using a single count and we can
optimize that with the hardware. The RMA working group is working on
an info key that will essentially do the same thing.
Note the above parameter won't help you with IB if you are using UCX
btl_uct_transports=dc_mlx5
btl=self,vader,uct
osc=^ucx
Though there may be a way to get osc/ucx to enable the same sort of
optimization. I don't know.
-Nathan

Post by Joseph Schuchart
All,
I am currently experimenting with MPI atomic operations and wanted to
share some interesting results I am observing. The numbers below are
measurements from both an IB-based cluster and our Cray XC40. The
```
if (rank == 1) {
uint64_t res, val;
for (size_t i = 0; i < NUM_REPS; ++i) {
MPI_Fetch_and_op(&val, &res, MPI_UINT32_T, 0, 0, MPI_SUM, win);
MPI_Win_flush(target, win);
}
}
MPI_Barrier(MPI_COMM_WORLD);
```
Only rank 1 performs atomic operations, rank 0 waits in a barrier (I
have tried to confirm that the operations are done in hardware by
letting rank 0 sleep for a while and ensuring that communication
progresses). Of particular interest for my use-case is fetch_op but I am
* Linux Cluster, IB QDR *
average of 100000 iterations
fetch_op: 4.323384us
compare_exchange: 2.035905us
accumulate: 4.326358us
get_accumulate: 4.334831us
fetch_op: 2.438080us
compare_exchange: 2.398836us
accumulate: 2.435378us
get_accumulate: 2.448347us
fetch_op: 6.819977us
compare_exchange: 4.551417us
accumulate: 6.807766us
get_accumulate: 6.817602us
fetch_op: 4.954860us
compare_exchange: 2.399373us
accumulate: 4.965702us
get_accumulate: 4.977876us
a) operations on 64bit operands generally seem to have lower latencies
than operations on 32bit
b) Using an exclusive lock leads to lower latencies
Overall, there is a factor of almost 3 between SharedLock+uint32_t and
ExclusiveLock+uint64_t for fetch_and_op, accumulate, and get_accumulate
(compare_exchange seems to be somewhat of an outlier).
* Cray XC40, Aries *
average of 100000 iterations
fetch_op: 2.011794us
compare_exchange: 1.740825us
accumulate: 1.795500us
get_accumulate: 1.985409us
fetch_op: 2.017172us
compare_exchange: 1.846202us
accumulate: 1.812578us
get_accumulate: 2.005541us
fetch_op: 5.380455us
compare_exchange: 5.164458us
accumulate: 5.230184us
get_accumulate: 5.399722us
fetch_op: 5.415230us
compare_exchange: 1.855840us
accumulate: 5.212632us
get_accumulate: 5.396110us
The difference between exclusive and shared lock is about the same as
with IB and the latencies for 32bit vs 64bit are roughly the same
(except for compare_exchange, it seems).
So my question is: is this to be expected? Is the higher latency when
using a shared lock caused by an internal lock being acquired because
the hardware operations are not actually atomic?
I'd be grateful for any insight on this.
Cheers,
Joseph
--
Dipl.-Inf. Joseph Schuchart
High Performance Computing Center Stuttgart (HLRS)
Nobelstr. 19
D-70569 Stuttgart
Tel.: [phone]
Fax: [phone]
_______________________________________________
users mailing list
https://lists.open-mpi.org/mailman/listinfo/users

_______________________________________________
users mailing list
https://lists.open-mpi.org/mailman/listinfo/users

_______________________________________________
users mailing list
https://lists.open-mpi.org/mailman/listinfo/users

_______________________________________________
users mailing list
https://lists.open-mpi.org/mailman/listinfo/users

_______________________________________________
users mailing list
https://lists.open-mpi.org/mailman/listinfo/users

[Nathan Hjelm via users]

Ok, then it sounds like a regression. I will try to track it down today or tomorrow.


-Nathan

On Nov 08, 2018, at 01:41 PM, Joseph Schuchart <***@hlrs.de> wrote:


Sorry for the delay, I wanted to make sure that I test the same version
on both Aries and IB: git master bbe5da4. I realized that I had
previously tested with 3.1.3 on the IB cluster, which ran fine. If I use
the same version I run into the same problem on both systems (with --mca
btl_openib_allow_ib true --mca osc_rdma_acc_single_intrinsic true). I
have not tried using UCX for this.

Joseph

On 11/8/18 1:20 PM, Nathan Hjelm via users wrote:

Quick scan of the program and it looks ok to me. I will dig deeper and
see if I can determine the underlying cause.


What Open MPI version are you using?


-Nathan


On Nov 08, 2018, at 11:10 AM, Joseph Schuchart <***@hlrs.de> wrote:


While using the mca parameter in a real application I noticed a strange
effect, which took me a while to figure out: It appears that on the
Aries network the accumulate operations are not atomic anymore. I am
attaching a test program that shows the problem: all but one processes
continuously increment a counter while rank 0 is continuously
subtracting a large value and adding it again, eventually checking for
the correct number of increments. Without the mca parameter the test at
the end succeeds as all increments are accounted for:


```
$ mpirun -n 16 -N 1 ./mpi_fetch_op_local_remote
result:15000
```


When setting the mca parameter the test fails with garbage in the result:


```
$ mpirun --mca osc_rdma_acc_single_intrinsic true -n 16 -N 1
./mpi_fetch_op_local_remote
result:25769849013
mpi_fetch_op_local_remote: mpi_fetch_op_local_remote.c:97: main:
Assertion `sum == 1000*(comm_size-1)' failed.
```


All processes perform only MPI_Fetch_and_op in combination with MPI_SUM
so I assume that the test in combination with the mca flag is correct. I
cannot reproduce this issue on our IB cluster.


Is that an issue in Open MPI or is there some problem in the test case
that I am missing?


Thanks in advance,
Joseph




On 11/6/18 1:15 PM, Joseph Schuchart wrote:
Thanks a lot for the quick reply, setting
osc_rdma_acc_single_intrinsic=true does the trick for both shared and
exclusive locks and brings it down to <2us per operation. I hope that
the info key will make it into the next version of the standard, I
certainly have use for it :)


Cheers,
Joseph


On 11/6/18 12:13 PM, Nathan Hjelm via users wrote:


All of this is completely expected. Due to the requirements of the
standard it is difficult to make use of network atomics even for
MPI_Compare_and_swap (MPI_Accumulate and MPI_Get_accumulate spoil the
party). If you want MPI_Fetch_and_op to be fast set this MCA parameter:


osc_rdma_acc_single_intrinsic=true


Shared lock is slower than an exclusive lock because there is an extra
lock step as part of the accumulate (it isn't needed if there is an
exclusive lock). When setting the above parameter you are telling the
implementation that you will only be using a single count and we can
optimize that with the hardware. The RMA working group is working on
an info key that will essentially do the same thing.


Note the above parameter won't help you with IB if you are using UCX
unless you set this (master only right now):


btl_uct_transports=dc_mlx5


btl=self,vader,uct


osc=^ucx




Though there may be a way to get osc/ucx to enable the same sort of
optimization. I don't know.




-Nathan




On Nov 06, 2018, at 09:38 AM, Joseph Schuchart <***@hlrs.de>
wrote:


All,


I am currently experimenting with MPI atomic operations and wanted to
share some interesting results I am observing. The numbers below are
measurements from both an IB-based cluster and our Cray XC40. The
benchmarks look like the following snippet:


```
if (rank == 1) {
uint64_t res, val;
for (size_t i = 0; i < NUM_REPS; ++i) {
MPI_Fetch_and_op(&val, &res, MPI_UINT32_T, 0, 0, MPI_SUM, win);
MPI_Win_flush(target, win);
}
}
MPI_Barrier(MPI_COMM_WORLD);
```


Only rank 1 performs atomic operations, rank 0 waits in a barrier (I
have tried to confirm that the operations are done in hardware by
letting rank 0 sleep for a while and ensuring that communication
progresses). Of particular interest for my use-case is fetch_op but
I am
including other operations here nevertheless:


* Linux Cluster, IB QDR *
average of 100000 iterations


Exclusive lock, MPI_UINT32_T:
fetch_op: 4.323384us
compare_exchange: 2.035905us
accumulate: 4.326358us
get_accumulate: 4.334831us


Exclusive lock, MPI_UINT64_T:
fetch_op: 2.438080us
compare_exchange: 2.398836us
accumulate: 2.435378us
get_accumulate: 2.448347us


Shared lock, MPI_UINT32_T:
fetch_op: 6.819977us
compare_exchange: 4.551417us
accumulate: 6.807766us
get_accumulate: 6.817602us


Shared lock, MPI_UINT64_T:
fetch_op: 4.954860us
compare_exchange: 2.399373us
accumulate: 4.965702us
get_accumulate: 4.977876us


There are two interesting observations:
a) operations on 64bit operands generally seem to have lower latencies
than operations on 32bit
b) Using an exclusive lock leads to lower latencies


Overall, there is a factor of almost 3 between SharedLock+uint32_t and
ExclusiveLock+uint64_t for fetch_and_op, accumulate, and get_accumulate
(compare_exchange seems to be somewhat of an outlier).


* Cray XC40, Aries *
average of 100000 iterations


Exclusive lock, MPI_UINT32_T:
fetch_op: 2.011794us
compare_exchange: 1.740825us
accumulate: 1.795500us
get_accumulate: 1.985409us


Exclusive lock, MPI_UINT64_T:
fetch_op: 2.017172us
compare_exchange: 1.846202us
accumulate: 1.812578us
get_accumulate: 2.005541us


Shared lock, MPI_UINT32_T:
fetch_op: 5.380455us
compare_exchange: 5.164458us
accumulate: 5.230184us
get_accumulate: 5.399722us


Shared lock, MPI_UINT64_T:
fetch_op: 5.415230us
compare_exchange: 1.855840us
accumulate: 5.212632us
get_accumulate: 5.396110us




The difference between exclusive and shared lock is about the same as
with IB and the latencies for 32bit vs 64bit are roughly the same
(except for compare_exchange, it seems).


So my question is: is this to be expected? Is the higher latency when
using a shared lock caused by an internal lock being acquired because
the hardware operations are not actually atomic?


I'd be grateful for any insight on this.


Cheers,
Joseph


--
Dipl.-Inf. Joseph Schuchart
High Performance Computing Center Stuttgart (HLRS)
Nobelstr. 19
D-70569 Stuttgart


Tel.: [phone]
Fax: [phone]
E-Mail: ***@hlrs.de <mailto:***@hlrs.de>
<mailto:***@hlrs.de>
_______________________________________________
users mailing list
***@lists.open-mpi.org <mailto:***@lists.open-mpi.org>
<mailto:***@lists.open-mpi.org>
https://lists.open-mpi.org/mailman/listinfo/users


_______________________________________________
users mailing list
***@lists.open-mpi.org <mailto:***@lists.open-mpi.org>
https://lists.open-mpi.org/mailman/listinfo/users


_______________________________________________
users mailing list
***@lists.open-mpi.org <mailto:***@lists.open-mpi.org>
https://lists.open-mpi.org/mailman/listinfo/users
_______________________________________________
users mailing list
***@lists.open-mpi.org <mailto:***@lists.open-mpi.org>
https://lists.open-mpi.org/mailman/listinfo/users