search

aws / aws-ofi-nccl

This is a plugin which lets EC2 developers use libfabric as network provider while running NCCL applications.
Apache License 2.0
150 stars 57 forks source link

register_mr_buffers:544 NCCL WARN NET/OFI Unable to register memory (type = 2) for device 0. RC: -22, Error: Invalid argument #584

Open visatish opened 2 months ago

visatish commented 2 months ago

Hi,

I'm trying to run a nccl allreduce benchmark on AWS EC2 and running into the following error:

register_mr_buffers:544 NCCL WARN NET/OFI Unable to register memory (type = 2) for device 0. RC: -22, Error: Invalid argument

Setup:

2x p4d.24xlarge

"Deep Learning AMI GPU PyTorch 2.1.0 (Ubuntu 20.04)" AMI

Relevant libs (note that I have installed the latest torch 2.4.1 & deps fresh):

Single EFA-enabled NIC (note that I know this instance type can support up to 4x, but I'm starting with 1):

(base) ubuntu@ip-172-31-36-110:~$ fi_info -p efa -t FI_EP_RDM
provider: efa
    fabric: efa
    domain: rdmap16s27-rdm
    version: 118.20
    type: FI_EP_RDM
    protocol: FI_PROTO_EFA
(base) ubuntu@ip-172-31-32-222:~$ fi_info --version
fi_info: 1.18.2amzn1.0
libfabric: 1.18.2amzn1.0
libfabric api: 1.18
(base) ubuntu@ip-172-31-36-110:~$ lspci -i efa
00:00.0 Host bridge: Intel Corporation 440FX - 82441FX PMC [Natoma]
00:01.0 ISA bridge: Intel Corporation 82371SB PIIX3 ISA [Natoma/Triton II]
00:01.3 Non-VGA unclassified device: Intel Corporation 82371AB/EB/MB PIIX4 ACPI (rev 08)
00:03.0 VGA compatible controller: Amazon.com, Inc. Device 1111
00:04.0 Non-Volatile memory controller: Amazon.com, Inc. Device 8061
10:00.0 Ethernet controller: Amazon.com, Inc. Elastic Network Adapter (ENA)
10:1b.0 Ethernet controller: Amazon.com, Inc. Elastic Fabric Adapter (EFA)
10:1c.0 3D controller: NVIDIA Corporation Device 20b0 (rev a1)
10:1d.0 3D controller: NVIDIA Corporation Device 20b0 (rev a1)
10:1e.0 Non-Volatile memory controller: Amazon.com, Inc. NVMe SSD Controller
10:1f.0 Non-Volatile memory controller: Amazon.com, Inc. NVMe SSD Controller
20:1c.0 3D controller: NVIDIA Corporation Device 20b0 (rev a1)
20:1d.0 3D controller: NVIDIA Corporation Device 20b0 (rev a1)
20:1e.0 Non-Volatile memory controller: Amazon.com, Inc. NVMe SSD Controller
20:1f.0 Non-Volatile memory controller: Amazon.com, Inc. NVMe SSD Controller
80:1a.0 Bridge: NVIDIA Corporation Device 1af1 (rev a1)
80:1b.0 Bridge: NVIDIA Corporation Device 1af1 (rev a1)
80:1c.0 Bridge: NVIDIA Corporation Device 1af1 (rev a1)
80:1d.0 Bridge: NVIDIA Corporation Device 1af1 (rev a1)
80:1e.0 Bridge: NVIDIA Corporation Device 1af1 (rev a1)
80:1f.0 Bridge: NVIDIA Corporation Device 1af1 (rev a1)
90:1c.0 3D controller: NVIDIA Corporation Device 20b0 (rev a1)
90:1d.0 3D controller: NVIDIA Corporation Device 20b0 (rev a1)
90:1e.0 Non-Volatile memory controller: Amazon.com, Inc. NVMe SSD Controller
90:1f.0 Non-Volatile memory controller: Amazon.com, Inc. NVMe SSD Controller
a0:1c.0 3D controller: NVIDIA Corporation Device 20b0 (rev a1)
a0:1d.0 3D controller: NVIDIA Corporation Device 20b0 (rev a1)
a0:1e.0 Non-Volatile memory controller: Amazon.com, Inc. NVMe SSD Controller
a0:1f.0 Non-Volatile memory controller: Amazon.com, Inc. NVMe SSD Controller

Cmd:

From https://github.com/stas00/ml-engineering.git:

cd ml-engineering/network/benchmarks
NCCL_DEBUG=INFO python -u -m torch.distributed.run --nproc_per_node 8 --nnodes 2 --rdzv_endpoint <head node addr>:8888 --rdzv_backend c10d --max_restarts 0 --role `hostname -s`: --tee 3 all_reduce_bench.py

Output:

nccl_out.txt

Note this particular portion:

(head, rank=0, pid=32220) [ip-172-31-33-151:1]:ip-172-31-33-151:38696:38818 [1] register_mr_buffers:544 NCCL WARN NET/OFI Unable to register memory (type = 2) for device 0. RC: -22, Error: Invalid argument
(head, rank=0, pid=32220) [ip-172-31-33-151:1]:ip-172-31-33-151:38696:38818 [1] NCCL INFO transport/net.cc:779 -> 2
(head, rank=0, pid=32220) [ip-172-31-33-151:1]:ip-172-31-33-151:38696:38818 [1] NCCL INFO misc/socket.cc:47 -> 3
(head, rank=0, pid=32220) [ip-172-31-33-151:1]:ip-172-31-33-151:38696:38818 [1] NCCL INFO misc/socket.cc:58 -> 3
(head, rank=0, pid=32220) [ip-172-31-33-151:1]:ip-172-31-33-151:38696:38818 [1] NCCL INFO misc/socket.cc:775 -> 3
(head, rank=0, pid=32220) [ip-172-31-33-151:1]:ip-172-31-33-151:38696:38818 [1] NCCL INFO proxy.cc:1384 -> 3
(head, rank=0, pid=32220) [ip-172-31-33-151:1]:
(head, rank=0, pid=32220) [ip-172-31-33-151:1]:ip-172-31-33-151:38696:38818 [1] proxy.cc:1533 NCCL WARN [Service thread] Error encountered progressing operation=Connect, res=3, closing connection
(head, rank=0, pid=32220) [ip-172-31-33-151:1]:
(head, rank=0, pid=32220) [ip-172-31-33-151:1]:ip-172-31-33-151:38696:38818 [1] proxy.cc:1567 NCCL WARN [Proxy Service 1] Failed to execute operation Connect from rank 1, retcode 3
(worker1, rank=1, pid=30588, ip=172.31.42.166) [ip-172-31-42-166:1]:
(worker1, rank=1, pid=30588, ip=172.31.42.166) [ip-172-31-42-166:1]:ip-172-31-42-166:30780:30883 [1] register_mr_buffers:544 NCCL WARN NET/OFI Unable to register memory (type = 2) for device 0. RC: -22, Error: Invalid argument
(worker1, rank=1, pid=30588, ip=172.31.42.166) [ip-172-31-42-166:1]:ip-172-31-42-166:30780:30883 [1] NCCL INFO transport/net.cc:779 -> 2
(worker1, rank=1, pid=30588, ip=172.31.42.166) [ip-172-31-42-166:1]:ip-172-31-42-166:30780:30872 [1] NCCL INFO transport/net.cc:304 -> 2
(worker1, rank=1, pid=30588, ip=172.31.42.166) [ip-172-31-42-166:1]:ip-172-31-42-166:30780:30872 [1] NCCL INFO transport.cc:165 -> 2
(worker1, rank=1, pid=30588, ip=172.31.42.166) [ip-172-31-42-166:1]:ip-172-31-42-166:30780:30872 [1] NCCL INFO init.cc:1222 -> 2
(worker1, rank=1, pid=30588, ip=172.31.42.166) [ip-172-31-42-166:1]:ip-172-31-42-166:30780:30872 [1] NCCL INFO init.cc:1501 -> 2
(worker1, rank=1, pid=30588, ip=172.31.42.166) [ip-172-31-42-166:1]:ip-172-31-42-166:30780:30872 [1] NCCL INFO group.cc:64 -> 2 [Async thread]
(worker1, rank=1, pid=30588, ip=172.31.42.166) [ip-172-31-42-166:1]:ip-172-31-42-166:30780:30780 [1] NCCL INFO group.cc:418 -> 2
(worker1, rank=1, pid=30588, ip=172.31.42.166) [ip-172-31-42-166:1]:ip-172-31-42-166:30780:30780 [1] NCCL INFO init.cc:1876 -> 2

I'm not quite sure what Error: Invalid argument could be - any help is appreciated. Thnx!

visatish commented 2 months ago

@bwbarrett I noticed you had helped with some related issues

AmedeoSapio commented 2 months ago

Hi, can you please try enabling all 4 EFAs?

visatish commented 2 months ago

@AmedeoSapio I was actually able to get it working with the native pytorch version in the AMI, i.e. conda activate pytorch:

(head, rank=0, pid=36870) [ip-172-31-40-103:0]:The average bandwidth of all_reduce with a 4.0GB payload (5 trials, 16 ranks):
(head, rank=0, pid=36870) [ip-172-31-40-103:0]: algbw: 11.135 GBps (89.1 Gbps)
(head, rank=0, pid=36870) [ip-172-31-40-103:0]: busbw: 20.878 GBps (167.0 Gbps)

I will try with 4 NICs, but presumably that will just increase bandwidth.

This hints that there is some incompatibility between aws-ofi-nccl and the latest torch + torch deps (I have updated the original issue to note that I was installing the latest fresh - i.e. pip install torch before running cmds).

rauteric commented 2 months ago

Hello. There is a known incompatibility between NCCL 2.19+ and Libfabric from EFA installers before 1.29. I'm guessing using the latest PyTorch will upgrade the NCCL version.

Workarounds are any of the following:

  1. Set FI_EFA_SET_CUDA_SYNC_MEMOPS=0 in the environment
  2. Downgrade to NCCL 2.18 (which it sounds like using native PyTorch will do)
  3. Upgrade to EFA installer 1.29 or greater (latest is 1.34)
visatish commented 2 months ago

Hi @rauteric, good to know! Is there any significant performance downside to (1) as that would be the least-invasive for our stack atm?

rauteric commented 2 months ago

Hi @rauteric, good to know! Is there any significant performance downside to (1) as that would be the least-invasive for our stack atm?

No, this setting merely prevents Libfabric from setting a property on a CUDA buffer (sync_memops) that is not needed for NCCL. It shouldn't have any performance impact.

visatish commented 2 months ago

Gotcha, confirmed that FI_EFA_SET_CUDA_SYNC_MEMOPS=0 works with the latest pytorch+NCCL stack in the original example.

Might be nice for future new users to maybe "pin" this in some fashion under "Known problems/limitations" in an easy-to-find place or have an up-to-date compatibility chart. But for now, guess it's indexed in this ticket :)

Thanks again for the help!

aws-nslick commented 2 months ago

For future searchers, if it's at all possible, please do prefer to update efa.ko and libfabric instead of relying on this environment variable -- this specific workaround doesn't come with a perf hit, but you are missing out on other performance improvements and bug fixes by using older versions, and you should update whenever you can.

sean-smith commented 1 month ago

@visatish we've documented a bunch of these efa/nccl related failure modes in awesome-distributed-training repo, i.e. https://github.com/aws-samples/awsome-distributed-training/issues/203