microsoft/mscclpp
1
0
Fork 0
mirror of https://github.com/microsoft/mscclpp.git synced 2026-05-12 17:26:04 +00:00
Code Issues Packages Projects Releases Wiki Activity
565 Commits 104 Branches 16 Tags
69938b47aadc59cd2c424dd8565ec3428fa7d222
Commit Graph

32 Commits

Author SHA1 Message Date
Olli Saarikivi
8cb63a7d1a Get rid of comm.setup() 2023-11-08 18:44:45 +00:00
Binyang2014
6f43282c1d Fix allreduce bug (#197) 
Fix allreduce correctness issue
 2023-10-18 23:16:57 +08:00
Changho Hwang
8c0f9e84d0 v0.3.0 (#171) 2023-10-11 22:35:54 +08:00
Changho Hwang
6c0ee72916 Construct ProxyChannel with shared pointers (#184) 2023-09-18 05:46:23 +00:00
Changho Hwang
d2f13f1e54 Fix #174 (#180) 
Added `extern "C"` based on another specification in
`/usr/local/cuda/include/crt/common_functions.h`.
 2023-09-15 06:44:41 +00:00
Binyang2014
952f2da9cc Improve single node allreduce performance (#169) 
Improve all reduce performance for single node.
New number:
|   n_ctx | size    |  target latency (us) | allreduce5 | allreduce6 |
|---------|---------|----------------|------------|------------|
|       1 | 24.0kB  |            7.7 |            |        7.23|
|       2 | 48.0kB  |            7.7 |            |        7.69|
|       4 | 96.0kB  |            8   |            |        8.34|
|       8 | 192.0kB |           12.6 |            |        9.75|
|      12 | 288.0kB |           13   |            |       11.34|
|      16 | 384.0kB |           13.3 |            |       12.99|
|     768 | 18.0MB  |          158.7 |       160.3|            |
|     896 | 21.0MB  |          184.5 |       183.8|            |
|    1024 | 24.0MB  |          209.5 |       207.5|            |
|    1152 | 27.0MB  |          234.3 |       231.9|            |
|    1280 | 30.0MB  |          260   |       255.6|            |
|    1408 | 33.0MB  |          284.9 |       278.7|            |
|    1536 | 36.0MB  |          310.3 |       302.0|            |
|    1664 | 39.0MB  |          336.2 |       325.3|            |
|    1792 | 42.0MB  |          361.4 |       348.8|            |
|    1920 | 45.0MB  |          384.6 |       372.2|            |
|    2048 | 48.0MB  |          409.1 |       395.4|            |

---------

Co-authored-by: Changho Hwang <changhohwang@microsoft.com>
 2023-09-13 14:30:08 +00:00
Binyang2014
097aa8843a Fix pytest unstable issue. (#170) 
- remove `#include <cstdint>` from `poll.hpp`. To make it only contains
device-side code
- Fix compilation issue, which will cause pytest fail randomly. Reuse
the compiled result for same kernel with different arguments
 2023-09-06 17:09:04 -07:00
Olli Saarikivi
828be48b21 Add Context and Endpoint classes to enable non-Communicator use-cases (#166) 
This PR implements and closes #137. The new `Endpoint` and `Context`
classes expose the connection establishing functionality from
`Communicator`, which now is only responsible for tying together the
bootstrapper with a context.

The largest breaking change here is that
`Communicator.connectOnSetup(...)` now returns the `Connection` wrapped
inside a `NonblockingFuture`. This is because with the way `Context` is
implemented a `Connection` is now fully initialized on construction.

Some smaller breaking API changes from this change are that
`RegisteredMemory` no longer has a `rank()` function (as there maybe no
concept of rank), and similarly `Connection` has no `remoteRank()` and
`tag()` functions. The latter are replaced by `remoteRankOf` and `tagOf`
functions in `Communicator`.

A new `EndpointConfig` class is introduced to avoid duplication of the
IB configuration parameters in the APIs of `Context` and `Communicator`.
The usual usage pattern of just passing in a `Transport` still works due
to an implicit conversion into `EndpointConfig`.

Miscellaneous changes:
-Cleans up how the PIMPL pattern is applied by making both the `Impl`
struct and the `pimpl_` pointers private for all relevant classes in the
core API.
-Enables ctest to be run from the build root directory.
 2023-09-06 13:10:04 +08:00
Saeed Maleki
8d1b984bed Change device handle interfaces & others (#142) 
* Changed device handle interfaces
* Changed proxy service interfaces
* Move device code into separate files
* Fixed FIFO polling issues
* Add configuration arguments in several interface functions

---------

Co-authored-by: Changho Hwang <changhohwang@microsoft.com>
Co-authored-by: Binyang Li <binyli@microsoft.com>
Co-authored-by: root <root@a100-saemal0.qxveptpukjsuthqvv514inp03c.gx.internal.cloudapp.net>
 2023-08-16 20:00:56 +08:00
Binyang2014
a58e2e9623 Make sure the semaphore not be released during the lifecycle of SmChannel (#131) 
Fix #126

 - Put `std::shared_ptr<SmDevice2DeviceSemaphore>` into the `SmChannel` 
 - add a `DeviceHandle` struct in `SmChannel`
 - add `DeviceHandle` template
 
Users need to write code like this to use channel in device side:
```
using DeviceHandle = mscclpp::DeviceHandle<T>;
__device__ DeviceHandle<mscclpp::SimpleProxyChannel> channel;
__device__ DeviceHandle<mscclpp::SmChannel> smChannel;
```

To cover a channel to deviceHandle, need to call this function:
`mscclpp::deviceHandle(SimpleProxyChannel or SmChannel)`

---------

Co-authored-by: Changho Hwang <changhohwang@microsoft.com>
 2023-07-20 12:18:22 +08:00
Saeed Maleki
e7d5e652df Python bindings (#125) 
Co-authored-by: Olli Saarikivi <olsaarik@microsoft.com>
Co-authored-by: Changho Hwang <changhohwang@microsoft.com>
Co-authored-by: Binyang Li <binyli@microsoft.com>
 2023-07-19 15:35:54 +08:00
Changho Hwang
1d71715d19 Separate mscclpp-test kernels (#122) 
Separate different kernel implementations in mscclpp-test to reduce the
number of registers required by the kernels.
 2023-07-10 10:11:20 -07:00
Changho Hwang
4114d65c60 Documents & minor updates (#119) 
Co-authored-by: Saeed Maleki <saemal@microsoft.com>
Co-authored-by: Binyang Li <binyli@microsoft.com>
 2023-07-07 17:35:05 +08:00
Changho Hwang
bb7b85a810 2-node AllReduce improvements (#118) 
* Added `get()` interfaces to `SmChannel`
* Improved 2-node (8 gpus/node) AllReduce: algbw 139GB/s for 1GB (kernel
3) and 99GB/s for 48MB (kernel 4)
* Fixed a FIFO perf bug
* Several fixes & validations in mscclpp-test

---------

Co-authored-by: Binyang Li <binyli@microsoft.com>
Co-authored-by: Saeed Maleki <saemal@microsoft.com>
 2023-07-07 07:05:46 +00:00
Changho Hwang
6ec585f3d8 Packet copy for IB (#109) 
* Extend channels to support LL with IB
* Rename classes and interfaces
 2023-06-28 10:39:31 -07:00
Changho Hwang
21eed722af Add license comments (#106) 2023-06-25 12:40:12 +08:00
Binyang2014
2640578b22 Add performance check for mscclpp-test (#110) 
- Add ndmv4 perf baseline
- change mscclpp-test to output perf number into a json file
- add python script to check the perf result with the baseline
 2023-06-21 07:42:53 +00:00
Changho Hwang
60b3dd5a61 Bug fixes & resolve warnings (#107) 
* Fix a bug in host hashing
* Fix a bug in `HostEpoch::wait()`
* Remove misc warnings
 2023-06-16 09:31:23 +00:00
Binyang2014
8410fcd8fc Fix allgather kernel 2 perf bug (#108) 
Fix #105
 2023-06-16 15:36:20 +08:00
Binyang2014
8efacae332 update pipeline (#103) 
Update Azure pipeline:
- Using mscclpp:base-cuda12.1 image for building and testing
- Add mp-ut tests for multi-nodes
 2023-06-14 20:14:57 +08:00
Changho Hwang
4d0b0a650f Remove vulnerable sscanf (#101) 2023-06-14 10:02:46 +08:00
Binyang2014
b1ce368656 Implement host offload algorithm for allgather (#84) 
Implement host offload algorithm for allgather
For 1n-8p
```
# Initializing MSCCL++
# Setting up the connection in MSCCL++
#
#                                    in-place                       out-of-place          
#       size         count     time   algbw   busbw  #wrong     time   algbw   busbw  #wrong
#        (B)    (elements)     (us)  (GB/s)  (GB/s)            (us)  (GB/s)  (GB/s)       
        1024            32    73.02    0.01    0.01      0

# Out of bounds values : 0 OK
#
```
For 2n-16p
```
# Initializing MSCCL++
# Setting up the connection in MSCCL++
#
#                                    in-place                       out-of-place          
#       size         count     time   algbw   busbw  #wrong     time   algbw   busbw  #wrong
#        (B)    (elements)     (us)  (GB/s)  (GB/s)            (us)  (GB/s)  (GB/s)       
        1024            16    90.30    0.01    0.01      0

# Out of bounds values : 0 OK
#
```
 2023-06-13 10:01:58 +00:00
Binyang2014
6ee4e80317 Create Azure pipeline for multi-node tests (#97) 
Create Azure pipeline to run mscclpp-test on multi-nodes
 2023-06-13 06:34:07 +00:00
Changho Hwang
76718e4015 Saemal/atomic signal (#96) 
* code complelete

* fix correctness issue

* Fix correctness issuee

* fix lint

* ass compile

* Fix build issue

* Fix runtime error

* Fix correctness issue

* Fix crash issue

* minor change

* Fix memory leak

* Fix review comments

* Finish allgather

* address comments

* load element to register first then store to remote address

* Finish allGather

* init

* Build connections

* allreduce_test works

* Bug fix

* Add CUDA flags

* Add packet copy (LL)

* Lint

* Set tmpPtr from constructors

* Lint

* Multiple blocks per peer

* Beautify

* Temporal ring reduce

* Ring reduce works correctly

* Overlapping

* Fix overlapping

* Improve vector sum

* figuring out how to use atomics

* working now

* wip

* Enhance LL AllReduce

* Support multiple blocks per peer

* Fix a ring reduce bug

* Fix a AllReduce kernel 2 bug

* Bug fix

* wip

* Make it compilable

* Lint

* Lint

* Minor changes

* Unit test to reproduce memory consistency bugs

* Unit test bug fixes

* Fixes

* Typo

* wip

* done with core

* wip

* wip

* compiles

* only the atomic is failing

* almost working

* all tests pass now

* clang-12

* More jailbreaks

* bug fix for common.cu

* adding stdint to concurrency.hpp

* Out-of-place for AllReduce kernel 2

* Optimize `sync()`

* Fix mp_unit_tests

* Init TestEngine with TestArgs

* Change common.cu into common.cc

* Cleanup common.hpp

* Lint

* fixes to the mscclpp-tests

* fixed common.cc

---------

Co-authored-by: Binyang Li <binyli@microsoft.com>
Co-authored-by: Saeed Maleki <saemal@microsoft.com>
 2023-06-12 21:38:06 -07:00
Changho Hwang
43de015f3f Add packet copy (LL) for AllReduce (#85) 2023-06-12 21:53:50 +08:00
Changho Hwang
5a4885ccbb Misc updates (#95) 2023-06-12 13:53:43 +08:00
Changho Hwang
798631bd52 Update unit tests (#81) 2023-06-08 09:58:05 +00:00
Changho Hwang
0c14a67ad2 [mscclpp-test] Add AllReduce and AllToAll tests (#83) 2023-06-07 10:58:47 +00:00
Changho Hwang
9cee6c4a74 Cleanup old files and functions (#86) 2023-06-01 17:34:57 +08:00
Olli Saarikivi
457c422791 Remove alloc.h and beef up cuda_utils.hpp (#82) 2023-05-24 08:34:18 +00:00
Binyang2014
216373eab2 Add allgather test to mscclpp-test (#78) 
Finish allGather

Co-authored-by: Changho Hwang <changhohwang@microsoft.com>
 2023-05-23 00:37:25 -07:00
Binyang2014
a3cf48cc5d Rewrite mscclpp-test with cpp style API (#77) 
- Rewrite mscclpp-test with cpp style API
- Add SM copy
- add new sendRecv test
 2023-05-19 14:14:19 +08:00