[UK] p2p performance

experimental/ukernel/Makefile

@@ -73,11 +73,12 @@ NVCCFLAGS  += -DUSE_REDIS_OOB --compiler-options "$(REDIS_CFLAGS)"
 LDFLAGS    += $(REDIS_LIBS)
 endif
 
-# UCCL collective support
+# RDMA core dependency (collective/rdma static library)
 UCCL_ROOT := $(abspath ../../)
 UCCL_RDMA_ROOT := $(UCCL_ROOT)/collective/rdma
 UCCL_RDMA_STATIC := $(UCCL_RDMA_ROOT)/librdma.a
-UCCL_CFLAGS := -I$(UCCL_ROOT) -I$(UCCL_ROOT)/collective/rdma -I$(UCCL_ROOT)/include
+UCCL_CFLAGS := -I$(UCCL_ROOT) -I$(UCCL_ROOT)/collective/rdma -I$(UCCL_ROOT)/include \
+               -I$(UCCL_ROOT)/p2p -I$(UCCL_ROOT)/p2p/include -I$(UCCL_ROOT)/p2p/rdma
 UCCL_LDFLAGS := $(UCCL_RDMA_STATIC) -libverbs -lrdmacm
 
 CXXFLAGS += $(UCCL_CFLAGS)
@@ -118,10 +119,10 @@ BENCH_BINS := $(BENCH_TRANSPORT_BIN) $(BENCH_GDRCOPY_BIN)
 STATIC_LIB  := libukernel.a
 SHARED_LIB  := libukernel.so
 
-# Build the core ukernel artifacts plus top-level benchmarks.
-all: $(STATIC_LIB) $(SHARED_LIB) bench
+# Build the core ukernel artifacts plus benchmarks.
+all: $(STATIC_LIB) $(SHARED_LIB) bench_all
 
-# Compilation rules - local includes must come before system/UCCL includes to avoid header conflicts
+# Compilation rules - local includes must come before system/RDMA headers to avoid conflicts
 %.o: %.cc
 	$(CXX) $(LOCAL_INCLUDES) $(CXXFLAGS) $(INCLUDES) -MMD -MP -c $< -o $@
 
@@ -131,34 +132,44 @@ all: $(STATIC_LIB) $(SHARED_LIB) bench
 %.o: %.cu
 	$(NVCC) $(LOCAL_INCLUDES) $(NVCCFLAGS) $(INCLUDES) -MMD -MP -c $< -o $@
 
-ensure-uccl-rdma:
+ensure-rdma-core:
 	@if [ ! -f "$(UCCL_RDMA_STATIC)" ]; then \
 		echo "[ukernel] building collective/rdma..."; \
 		$(MAKE) -C "$(UCCL_RDMA_ROOT)" build; \
 	fi
 
+# Backward-compatible alias.
+ensure-uccl-rdma: ensure-rdma-core
+
 # Static library
 $(STATIC_LIB): $(OBJ_CC) $(OBJ_CU)
 	ar rcs $@ $^
 
 # Shared library
-$(SHARED_LIB): $(OBJ_CC) $(OBJ_CU) | ensure-uccl-rdma
+$(SHARED_LIB): $(OBJ_CC) $(OBJ_CU) | ensure-rdma-core
 	$(CXX) -shared $(CXXFLAGS) $(INCLUDES) $^ $(LDFLAGS) -o $@
 
 # Benchmark build rules
 $(BENCH_TRANSPORT_OBJ): $(BENCH_TRANSPORT_SRC)
 	$(CXX) $(LOCAL_INCLUDES) $(CXXFLAGS) $(INCLUDES) -MMD -MP -c $< -o $@
 
-$(BENCH_TRANSPORT_BIN): $(BENCH_TRANSPORT_OBJ) $(STATIC_LIB) | ensure-uccl-rdma
+$(BENCH_TRANSPORT_BIN): $(BENCH_TRANSPORT_OBJ) $(STATIC_LIB) | ensure-rdma-core
 	$(CXX) $(LOCAL_INCLUDES) $(CXXFLAGS) $(INCLUDES) $< $(STATIC_LIB) $(LDFLAGS) -o $@
 
 $(BENCH_GDRCOPY_OBJ): $(BENCH_GDRCOPY_SRC)
 	$(CXX) $(BENCH_GDRCOPY_CXXFLAGS) $(BENCH_GDRCOPY_INCLUDES) -MMD -MP -c $< -o $@
 
-$(BENCH_GDRCOPY_BIN): $(BENCH_GDRCOPY_OBJ) | ensure-uccl-rdma
+$(BENCH_GDRCOPY_BIN): $(BENCH_GDRCOPY_OBJ) | ensure-rdma-core
 	$(CXX) $< -o $@ $(BENCH_GDRCOPY_LIBS)
 
-bench: $(BENCH_BINS)
+bench_all: $(BENCH_BINS)
+
+# Backward-compatible alias.
+bench: bench_all
+
+transport_bench: bench_transport
+
+bench_gdrcopy: $(BENCH_GDRCOPY_BIN)
 
 device_test:
 	$(MAKE) -C src/device test $(SUBMAKE_SM)
@@ -196,6 +207,4 @@ clean:
 # src/device/persistent_kernel_ops.o: src/device/persistent_kernel_ops.cu
 # 	$(NVCC) $(NVCCFLAGS) -Iinclude -Isrc/transport -Isrc/device -Isrc/device/fifo -MMD -MP -c $< -o $@
 
-.PHONY: all clean bench transport_bench device_test ccl_test transport_test transport_suite device_bench ensure-uccl-rdma
-
-transport_bench: $(BENCH_TRANSPORT_BIN)
+.PHONY: all clean bench bench_all bench_transport bench_gdrcopy transport_bench device_test ccl_test transport_test transport_suite device_bench ensure-rdma-core ensure-uccl-rdma

experimental/ukernel/README.md

@@ -1,36 +1,19 @@
 # UKernel
 
-Minimal build and test entry points for `experimental/ukernel`.
+`experimental/ukernel` is a layered runtime with three core modules:
 
-## Prerequisites
-
-- CUDA toolchain for NVIDIA builds
-- RDMA / verbs dependencies used by `transport` and `ccl`
-- system-installed GDRCopy (`gdrapi.h` + `libgdrapi`) for `device` and `ccl`
-- `torchrun` available for CCL multiprocess integration tests
-
-## Quick Install GDRCopy (System)
-
-`experimental/ukernel` no longer builds `thirdparty/gdrcopy`.  
-Please install GDRCopy from NVIDIA upstream on your machine first.
-
-Ubuntu quick path (source build):
+- `transport`: peer bootstrap, path selection, async send/recv, progress
+- `device`: device-side FIFO/task/worker execution
+- `ccl`: collective planning + executor on top of transport/device
 
-```bash
-sudo apt-get update
-sudo apt-get install -y build-essential dkms linux-headers-$(uname -r) libelf-dev
-git clone https://github.com/NVIDIA/gdrcopy.git
-cd gdrcopy
-make CUDA=/usr/local/cuda
-sudo make CUDA=/usr/local/cuda prefix=/usr/local install
-```
+## Prerequisites
 
-Optional: if `libgdrapi.so` is not in the default linker path, pass `GDRCOPY_LIBDIR` when building:
+- CUDA toolchain (NVIDIA build path)
+- RDMA / verbs dependencies (`libibverbs`, `librdmacm`)
+- system-installed GDRCopy (`gdrapi.h` + `libgdrapi.so`)
+- `torchrun` for multiprocess CCL/Python integration tests
 
-```bash
-cd experimental/ukernel
-make GDRCOPY_LIBDIR=/usr/local/lib
-```
+If `libgdrapi.so` is outside default linker paths, pass `GDRCOPY_LIBDIR`.
 
 ## Build
 
@@ -42,204 +25,87 @@ make clean -f Makefile
 make -j$(nproc) -f Makefile
 ```
 
-AMD / ROCm:
+ROCm:
 
 ```bash
 cd experimental/ukernel
 make clean -f Makefile.rocm
 make -j$(nproc) -f Makefile.rocm
 ```
 
-Common overrides:
+Common override example:
 
 ```bash
 make -f Makefile CUDA_PATH=/usr/local/cuda CONDA_LIB_HOME=/usr/lib SM=80
 ```
 
-## Test
+## Common Targets
 
-Run all transport tests:
+Tests:
 
 ```bash
-cd experimental/ukernel
 make transport_test
+make device_test SM=80
+make ccl_test SM=80
 ```
 
-Build transport benchmark:
-
-```bash
-cd experimental/ukernel
-make transport_bench
-```
-
-Manual two-process transport check:
-
-```bash
-cd experimental/ukernel/src/transport
-make test-integration
-./test_transport_integration communicator --role=server --case=exchange --exchanger-port 16979
-./test_transport_integration communicator --role=client --case=exchange --exchanger-ip 127.0.0.1 --exchanger-port 16979
-```
-
-Run all device tests:
+Benchmarks:
 
 ```bash
-cd experimental/ukernel
-make device_test SM=80
+make bench_transport
+make bench_gdrcopy
+make bench_all
 ```
 
-Run all CCL tests:
+Compatibility aliases:
 
 ```bash
-cd experimental/ukernel
-make ccl_test SM=80
+make transport_bench   # alias of bench_transport
+make bench             # alias of bench_all
 ```
 
 ## Python Binding
 
-`experimental/ukernel/py` contains a `torch`-based Python extension that wraps
-the `ccl` executor behind a persistent `ProcessGroup` object. The binding takes
-CUDA `torch.Tensor` inputs directly and runs `allreduce` / `alltoall` through
-the existing `transport + ccl + device` stack.
-
-Build the extension in place:
+Build extension:
 
 ```bash
 cd experimental/ukernel/py
 python setup.py build_ext --inplace
 ```
 
-Run Python tests (requires 2+ GPUs):
+Run common tests:
 
 ```bash
-cd experimental/ukernel/py
 CUDA_VISIBLE_DEVICES=6,7 torchrun --nproc_per_node=2 test_collective.py
-CUDA_VISIBLE_DEVICES=0,6,7 torchrun --nproc_per_node=3 test_collective.py  # for 3-rank tests
 CUDA_VISIBLE_DEVICES=6,7 torchrun --nproc_per_node=2 test_p2p.py
+```
+
+Run p2p benchmark flavors:
 
-# rdma battle
+```bash
+# Collective/rdma path
+# UCCL_DEBUG_VLOG_LEVEL=1 : check fast path
 UK_P2P_TRANSPORT=uccl UCCL_P2P_MODE=rdma NCCL_P2P_DISABLE=1 NCCL_SHM_DISABLE=1 NCCL_IB_DISABLE=0 NCCL_IB_HCA=mlx5_0 NCCL_DEBUG=INFO CUDA_VISIBLE_DEVICES=6,7 torchrun --nproc_per_node=2 bench_p2p.py
-        Size |  ukernel (ms) |  ukernel (GB/s) |   UCCL (ms) |   UCCL (GB/s) |   NCCL (ms) |   NCCL (GB/s)
---------------------------------------------------------------------------------------------------------------------------------
-      1024 B |         0.063 |            0.03 |       0.316 |          0.01 |       0.096 |          0.02
-      4096 B |         0.093 |            0.09 |       0.306 |          0.03 |       0.099 |          0.08
-     16384 B |         0.199 |            0.16 |       0.366 |          0.09 |       0.103 |          0.32
-     65536 B |         0.249 |            0.53 |       0.431 |          0.30 |       0.159 |          0.82
-    262144 B |         0.437 |            1.20 |       0.687 |          0.76 |       0.385 |          1.36
-   1048576 B |         1.238 |            1.69 |       1.846 |          1.14 |       1.244 |          1.69
-   4194304 B |         5.232 |            1.60 |       5.699 |          1.47 |       4.480 |          1.87
-  16777216 B |        19.853 |            1.69 |      20.715 |          1.62 |      17.434 |          1.92
-  67108864 B |        79.071 |            1.70 |      78.673 |          1.71 |      69.287 |          1.94
- 268435456 B |       319.966 |            1.68 |     319.911 |          1.68 |     275.867 |          1.95
-
-        Size |  ukernel (ms) |  ukernel (GB/s) |   UCCL (ms) |   UCCL (GB/s) |   NCCL (ms) |   NCCL (GB/s)
---------------------------------------------------------------------------------------------------------------------------------
-      1024 B |         0.074 |            0.03 |       0.711 |          0.00 |       0.097 |          0.02
-      4096 B |         0.086 |            0.10 |       0.316 |          0.03 |       0.098 |          0.08
-     16384 B |         0.196 |            0.17 |       0.240 |          0.14 |       0.099 |          0.33
-     65536 B |         0.234 |            0.56 |       0.300 |          0.44 |       0.150 |          0.88
-    262144 B |         0.416 |            1.26 |       0.631 |          0.83 |       0.372 |          1.41
-   1048576 B |         1.272 |            1.65 |       1.388 |          1.51 |       1.227 |          1.71
-   4194304 B |         5.083 |            1.65 |       5.726 |          1.46 |       4.475 |          1.87
-  16777216 B |        20.159 |            1.66 |      24.035 |          1.40 |      17.449 |          1.92
-  67108864 B |        79.323 |            1.69 |      78.945 |          1.70 |      69.256 |          1.94
- 268435456 B |       319.898 |            1.68 |     315.428 |          1.70 |     275.951 |          1.95
-
-        Size |  ukernel (ms) |  ukernel (GB/s) |   UCCL (ms) |   UCCL (GB/s) |   NCCL (ms) |   NCCL (GB/s)
---------------------------------------------------------------------------------------------------------------------------------
-      1024 B |         0.086 |            0.02 |       0.140 |          0.01 |       0.137 |          0.01
-      4096 B |         0.090 |            0.09 |       0.121 |          0.07 |       0.132 |          0.06
-     16384 B |         0.180 |            0.18 |       0.149 |          0.22 |       0.133 |          0.25
-     65536 B |         0.246 |            0.53 |       0.205 |          0.64 |       0.151 |          0.87
-    262144 B |         0.410 |            1.28 |       0.410 |          1.28 |       0.374 |          1.40
-   1048576 B |         1.268 |            1.65 |       1.243 |          1.69 |       1.225 |          1.71
-   4194304 B |         5.050 |            1.66 |       5.033 |          1.67 |       4.437 |          1.89
-  16777216 B |        19.814 |            1.69 |      19.923 |          1.68 |      17.447 |          1.92
-  67108864 B |        78.855 |            1.70 |      78.982 |          1.70 |      69.237 |          1.94
- 268435456 B |       319.691 |            1.68 |     315.106 |          1.70 |     276.135 |          1.94
-
-# ipc battle
-UK_P2P_TRANSPORT=ipc UCCL_P2P_MODE=ipc CUDA_VISIBLE_DEVICES=6,7 torchrun --nproc_per_node=2 bench_p2p.py
-         Size |  ukernel (ms) |  ukernel (GB/s) |   UCCL (ms) |   UCCL (GB/s) |   NCCL (ms) |   NCCL (GB/s)
---------------------------------------------------------------------------------------------------------------------------------
-      1024 B |         3.318 |            0.00 |       0.042 |          0.05 |       0.091 |          0.02
-      4096 B |         3.470 |            0.00 |       0.042 |          0.20 |       0.090 |          0.09
-     16384 B |         3.339 |            0.01 |       0.064 |          0.51 |       0.087 |          0.37
-     65536 B |         3.410 |            0.04 |       0.078 |          1.68 |       0.098 |          1.34
-    262144 B |         3.439 |            0.15 |       0.185 |          2.84 |       0.088 |          5.96
-   1048576 B |         3.602 |            0.58 |       0.566 |          3.70 |       0.102 |         20.66
-   4194304 B |         3.679 |            2.28 |       1.374 |          6.10 |       0.250 |         33.50
-  16777216 B |         4.104 |            8.18 |       4.683 |          7.17 |       0.799 |         41.99
-  67108864 B |         6.058 |           22.15 |      17.865 |          7.51 |       2.758 |         48.66
- 268435456 B |        13.548 |           39.63 |      70.594 |          7.61 |      10.814 |         49.64
-
-# tcp
-UK_P2P_TRANSPORT=tcp UCCL_P2P_MODE=ipc CUDA_VISIBLE_DEVICES=6,7 torchrun --nproc_per_node=2 bench_p2p.py
-         Size |  ukernel (ms) |  ukernel (GB/s)
-------------------------------------------------------
-      1024 B |         3.318 |            0.00
-      4096 B |         3.470 |            0.00
-     16384 B |         3.339 |            0.01
-     65536 B |         3.410 |            0.04
-    262144 B |         3.439 |            0.15
-   1048576 B |         3.602 |            0.58
-   4194304 B |         3.679 |            2.28
-  16777216 B |         4.104 |            8.18
-  67108864 B |         6.058 |           22.15
- 268435456 B |        13.548 |           39.63
-```
 
-Minimal usage under `torchrun`:
-
-```python
-import os
-import torch
-from ukernel_ccl import ProcessGroup
-
-rank = int(os.environ["RANK"])
-world = int(os.environ["WORLD_SIZE"])
-local_rank = int(os.environ.get("LOCAL_RANK", rank))
-
-torch.cuda.set_device(local_rank)
-pg = ProcessGroup(
-    rank=rank,
-    world_size=world,
-    gpu_id=local_rank,
-    exchanger_ip=os.environ.get("MASTER_ADDR", "127.0.0.1"),
-    exchanger_port=int(os.environ.get("MASTER_PORT", "29500")),
-    transport="auto",
-)
-
-x = torch.randn(1024 * world + 1, device="cuda", dtype=torch.float32)
-pg.allreduce(x, tile_bytes=65536, num_flows=2)
-
-y = torch.randn(1024 * world, device="cuda", dtype=torch.float32)
-pg.alltoall(y, tile_bytes=65536, num_flows=2)
-
-send = torch.randn(13, device="cuda", dtype=torch.float32)
-recv = torch.empty(13, device="cuda", dtype=torch.float32)
-dist.all_to_all_single(
-    recv,
-    send,
-    output_split_sizes=[4, 5, 4],
-    input_split_sizes=[4, 5, 4],
-    group=pg,
-    tile_bytes=65536,
-    num_flows=2,
-)
+# p2p/rdma path
+UK_P2P_TRANSPORT=rdma UCCL_P2P_MODE=rdma NCCL_P2P_DISABLE=1 NCCL_SHM_DISABLE=1 NCCL_IB_DISABLE=0 NCCL_IB_HCA=mlx5_0 NCCL_DEBUG=INFO CUDA_VISIBLE_DEVICES=6,7 torchrun --nproc_per_node=2 bench_p2p.py
+
+# IPC path
+UK_P2P_TRANSPORT=ipc UCCL_P2P_MODE=ipc NCCL_P2P_DISABLE=1 NCCL_SHM_DISABLE=1 NCCL_IB_DISABLE=0 NCCL_IB_HCA=mlx5_0 NCCL_DEBUG=INFO CUDA_VISIBLE_DEVICES=6,7 torchrun --nproc_per_node=2 bench_p2p.py
+
+# TCP path
+UK_P2P_TRANSPORT=tcp UCCL_P2P_MODE=ipc NCCL_P2P_DISABLE=1 NCCL_SHM_DISABLE=1 NCCL_IB_DISABLE=0 NCCL_IB_HCA=mlx5_0 NCCL_DEBUG=INFO CUDA_VISIBLE_DEVICES=6,7 torchrun --nproc_per_node=2 bench_p2p.py
 ```
 
-Current Python binding constraints:
+Naming note:
 
-- collective payload tensors must be CUDA and contiguous
-- `allreduce` supports non-divisible element counts
-- equal-split and variable-split `all_to_all_single` are both supported
-- variable-split `all_to_all_single` requires explicit
-  `input_split_sizes/output_split_sizes` whose sums match local tensor numel
+- `UK_P2P_TRANSPORT=uccl` is a compatibility-facing logical selector.
+- In ukernel transport, this selector currently maps to the RDMA adapter backend.
+- `UCCL_P2P_MODE` controls standalone UCCL behavior and is independent of ukernel adapter naming.
 
-## Modules
+## Module Docs
 
-- [`src/transport/README.md`](/Users/jacelau/code/opencode/uccl/experimental/ukernel/src/transport/README.md)
-- [`src/device/README.md`](/Users/jacelau/code/opencode/uccl/experimental/ukernel/src/device/README.md)
-- [`src/ccl/README.md`](/Users/jacelau/code/opencode/uccl/experimental/ukernel/src/ccl/README.md)
-- [`benchmarks/README.md`](/Users/jacelau/code/opencode/uccl/experimental/ukernel/benchmarks/README.md)
+- [transport README](/Users/jacelau/code/opencode/uccl/experimental/ukernel/src/transport/README.md)
+- [device README](/Users/jacelau/code/opencode/uccl/experimental/ukernel/src/device/README.md)
+- [ccl README](/Users/jacelau/code/opencode/uccl/experimental/ukernel/src/ccl/README.md)
+- [benchmarks README](/Users/jacelau/code/opencode/uccl/experimental/ukernel/benchmarks/README.md)