Merge pull request #61 from yuhc/fix_legacy_manager

Port legacy manager
utcs-scea · Jan 25, 2021 · 24589ac · 24589ac
2 parents 996a434 + a97d936
commit 24589ac
Show file tree

Hide file tree

Showing 19 changed files with 2,739 additions and 52 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -41,6 +41,7 @@ set(AVA_GEN_DEMO_SPEC OFF CACHE BOOL "Enable demo specification")
 
 set(AVA_MANAGER_GALVANIC OFF CACHE BOOL "Build Galvanic manager")
 set(AVA_MANAGER_KATANA OFF CACHE BOOL  "Build Katana manager")
+set(AVA_MANAGER_LEGACY OFF CACHE BOOL  "Build legacy (r1.0) manager")
 set(AVA_MANAGER_DEMO OFF CACHE BOOL  "Build demo manager")
 
 message(STATUS "Build TensorFlow CUDA specification:   ${AVA_GEN_TF_SPEC}")
@@ -58,6 +59,7 @@ message(STATUS "Build test (libtrivial) specification: ${AVA_GEN_TEST_SPEC}")
 message(STATUS "Build demo specification:              ${AVA_GEN_DEMO_SPEC}")
 message(STATUS "Build Galvanic manager:                ${AVA_MANAGER_GALVANIC}")
 message(STATUS "Build Katana manager:                  ${AVA_MANAGER_KATANA}")
+message(STATUS "Build legacy (r1.0) manager:           ${AVA_MANAGER_LEGACY}")
 message(STATUS "Build demo manager:                    ${AVA_MANAGER_DEMO}")
 
 set(AVA_INSTALL_DIR "${CMAKE_CURRENT_BINARY_DIR}/install" CACHE PATH "Installation path of AvA")
@@ -136,6 +138,7 @@ ExternalProject_Add(ava-manager
   CMAKE_CACHE_ARGS
         -DAVA_MANAGER_GALVANIC:BOOL=${AVA_MANAGER_GALVANIC}
         -DAVA_MANAGER_KATANA:BOOL=${AVA_MANAGER_KATANA}
+        -DAVA_MANAGER_LEGACY:BOOL=${AVA_MANAGER_LEGACY}
         -DAVA_MANAGER_DEMO:BOOL=${AVA_MANAGER_DEMO}
         -DCMAKE_TOOLCHAIN_FILE:STRING=${CMAKE_TOOLCHAIN_FILE}
         -DCMAKE_INSTALL_PREFIX:PATH=${AVA_INSTALL_DIR}

diff --git a/cava/nightwatch/generator/c/cmakelists.py b/cava/nightwatch/generator/c/cmakelists.py
@@ -14,7 +14,7 @@ def source(api: API, errors):
 
 list(APPEND CMAKE_MODULE_PATH "${{CMAKE_CURRENT_BINARY_DIR}}/../..")
 
-set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD 14)
 
 set(c_flags {api.cflags})
 set(cxx_flags {api.cxxflags})
@@ -37,7 +37,8 @@ def source(api: API, errors):
 find_package(Boost REQUIRED COMPONENTS system)
 find_library(Config++ NAMES libconfig++ config++ REQUIRED)
 
-###### Set generated files ######
+set(protobuf_MODULE_COMPATIBLE TRUE)
+find_package(Protobuf REQUIRED QUIET)
 
 ###### Compile ######
 
@@ -95,12 +96,14 @@ def source(api: API, errors):
   ${{CMAKE_SOURCE_DIR}}/../../common/cmd_channel_socket_utilities.cpp
   ${{CMAKE_SOURCE_DIR}}/../../common/cmd_channel_socket_tcp.cpp
   ${{CMAKE_SOURCE_DIR}}/../../common/cmd_channel_socket_vsock.cpp
+  ${{CMAKE_SOURCE_DIR}}/../../proto/manager_service.proto.cpp
 )
 target_link_libraries(guestlib
   ${{GLIB2_LIBRARIES}}
   ${{Boost_LIBRARIES}}
   Threads::Threads
   ${{Config++}}
+  ${{Protobuf_LIBRARIES}}
 )
 target_compile_options(guestlib
   PUBLIC -fvisibility=hidden

diff --git a/cava/samples/onnxruntime/onnx_dump.c b/cava/samples/onnxruntime/onnx_dump.c
@@ -7557,8 +7557,13 @@ cudnnGetConvolutionBackwardDataWorkspaceSize(cudnnHandle_t handle,
                                              cudnnConvolutionBwdDataAlgo_t algo,
                                              size_t *sizeInBytes)
 {
-    fprintf(stderr, "%s is not implemented\n", __func__);
-    abort();
+   ava_argument(wDesc) ava_handle;
+   ava_argument(dyDesc) ava_handle;
+   ava_argument(convDesc) ava_handle;
+   ava_argument(dxDesc) ava_handle;
+   ava_argument(sizeInBytes) {
+     ava_out; ava_buffer(1);
+   }
 }
 
 cudnnStatus_t CUDNNWINAPI
@@ -7576,8 +7581,26 @@ cudnnConvolutionBackwardData(cudnnHandle_t handle,
                              const cudnnTensorDescriptor_t dxDesc,
                              void *dx)
 {
-    fprintf(stderr, "%s is not implemented\n", __func__);
-    abort();
+   ava_async;
+   ava_argument(handle) ava_handle;
+   ava_argument(alpha) {
+      ava_type_cast(const double *);
+      ava_in; ava_buffer(1);
+   }
+   ava_argument(beta) {
+      ava_type_cast(const double *);
+      ava_in; ava_buffer(1);
+   }
+   ava_argument(dx) ava_opaque;
+   ava_argument(dxDesc) ava_handle;
+   ava_argument(wDesc) ava_handle;
+   ava_argument(w) ava_opaque;
+   ava_argument(dyDesc) ava_handle;
+   ava_argument(dy) ava_opaque;
+   ava_argument(convDesc) ava_handle;
+   ava_argument(workSpace) {
+     ava_in; ava_buffer(workSpaceSizeInBytes);
+   }
 }
 
 cudnnStatus_t CUDNNWINAPI

diff --git a/cava/samples/onnxruntime/onnx_opt.c b/cava/samples/onnxruntime/onnx_opt.c
@@ -8169,8 +8169,13 @@ cudnnGetConvolutionBackwardDataWorkspaceSize(cudnnHandle_t handle,
                                              cudnnConvolutionBwdDataAlgo_t algo,
                                              size_t *sizeInBytes)
 {
-    fprintf(stderr, "%s is not implemented\n", __func__);
-    abort();
+   ava_argument(wDesc) ava_handle;
+   ava_argument(dyDesc) ava_handle;
+   ava_argument(convDesc) ava_handle;
+   ava_argument(dxDesc) ava_handle;
+   ava_argument(sizeInBytes) {
+     ava_out; ava_buffer(1);
+   }
 }
 
 cudnnStatus_t CUDNNWINAPI
@@ -8188,8 +8193,26 @@ cudnnConvolutionBackwardData(cudnnHandle_t handle,
                              const cudnnTensorDescriptor_t dxDesc,
                              void *dx)
 {
-    fprintf(stderr, "%s is not implemented\n", __func__);
-    abort();
+   ava_async;
+   ava_argument(handle) ava_handle;
+   ava_argument(alpha) {
+      ava_type_cast(const double *);
+      ava_in; ava_buffer(1);
+   }
+   ava_argument(beta) {
+      ava_type_cast(const double *);
+      ava_in; ava_buffer(1);
+   }
+   ava_argument(dx) ava_opaque;
+   ava_argument(dxDesc) ava_handle;
+   ava_argument(wDesc) ava_handle;
+   ava_argument(w) ava_opaque;
+   ava_argument(dyDesc) ava_handle;
+   ava_argument(dy) ava_opaque;
+   ava_argument(convDesc) ava_handle;
+   ava_argument(workSpace) {
+     ava_in; ava_buffer(workSpaceSizeInBytes);
+   }
 }
 
 cudnnStatus_t CUDNNWINAPI

diff --git a/cava/samples/tensorflow/CMakeLists.txt b/cava/samples/tensorflow/CMakeLists.txt
@@ -32,8 +32,12 @@ ExternalProject_Add_Step(cava tf_dump-link
           mkdir -p ${tf_dump_install}/lib &&
           ln -f -s ${CMAKE_CURRENT_SOURCE_DIR}/headers ${CMAKE_CURRENT_BINARY_DIR}/headers        &&
           ln -f -s ${tf_dump_install}/lib/libguestlib.so ${tf_dump_install}/lib/libcuda.so.1      &&
+          ln -f -s ${tf_dump_install}/lib/libguestlib.so ${tf_dump_install}/lib/libcuda.so        &&
           ln -f -s ${tf_dump_install}/lib/libguestlib.so ${tf_dump_install}/lib/libcublas.so.10   &&
           ln -f -s ${tf_dump_install}/lib/libguestlib.so ${tf_dump_install}/lib/libcublasLt.so.10 &&
+          ln -f -s ${tf_dump_install}/lib/libguestlib.so ${tf_dump_install}/lib/libcudart.so.10.0 &&
+          ln -f -s ${tf_dump_install}/lib/libguestlib.so ${tf_dump_install}/lib/libcudart.so.10.1 &&
+          ln -f -s ${tf_dump_install}/lib/libguestlib.so ${tf_dump_install}/lib/libcudart.so.10.2 &&
           ln -f -s ${tf_dump_install}/lib/libguestlib.so ${tf_dump_install}/lib/libcudart.so.10   &&
           ln -f -s ${tf_dump_install}/lib/libguestlib.so ${tf_dump_install}/lib/libcudnn.so.7     &&
           ln -f -s ${tf_dump_install}/lib/libguestlib.so ${tf_dump_install}/lib/libcufft.so.10    &&
@@ -75,11 +79,17 @@ ExternalProject_Add_Step(cava tf_opt-link
           mkdir -p ${tf_opt_install}/lib &&
           ln -f -s ${CMAKE_CURRENT_SOURCE_DIR}/headers ${CMAKE_CURRENT_BINARY_DIR}/headers      &&
           ln -f -s ${tf_opt_install}/lib/libguestlib.so ${tf_opt_install}/lib/libcuda.so.1      &&
+          ln -f -s ${tf_opt_install}/lib/libguestlib.so ${tf_opt_install}/lib/libcuda.so        &&
           ln -f -s ${tf_opt_install}/lib/libguestlib.so ${tf_opt_install}/lib/libcublas.so.10   &&
           ln -f -s ${tf_opt_install}/lib/libguestlib.so ${tf_opt_install}/lib/libcublasLt.so.10 &&
+          ln -f -s ${tf_opt_install}/lib/libguestlib.so ${tf_opt_install}/lib/libcudart.so.10.0 &&
+          ln -f -s ${tf_opt_install}/lib/libguestlib.so ${tf_opt_install}/lib/libcudart.so.10.1 &&
+          ln -f -s ${tf_opt_install}/lib/libguestlib.so ${tf_opt_install}/lib/libcudart.so.10.2 &&
           ln -f -s ${tf_opt_install}/lib/libguestlib.so ${tf_opt_install}/lib/libcudart.so.10   &&
+          ln -f -s ${tf_opt_install}/lib/libguestlib.so ${tf_opt_install}/lib/libcudart.so      &&
           ln -f -s ${tf_opt_install}/lib/libguestlib.so ${tf_opt_install}/lib/libcudnn.so.7     &&
           ln -f -s ${tf_opt_install}/lib/libguestlib.so ${tf_opt_install}/lib/libcufft.so.10    &&
+          ln -f -s ${tf_opt_install}/lib/libguestlib.so ${tf_opt_install}/lib/libcufft.so.10    &&
           ln -f -s ${tf_opt_install}/lib/libguestlib.so ${tf_opt_install}/lib/libcurand.so.10   &&
           ln -f -s ${tf_opt_install}/lib/libguestlib.so ${tf_opt_install}/lib/libcusolver.so.10 &&
           ln -f -s ${tf_opt_install}/lib/libguestlib.so ${tf_opt_install}/lib/libcusparse.so.10

diff --git a/cava/samples/tensorflow/README.md b/cava/samples/tensorflow/README.md
@@ -1 +1,43 @@
+Install TensorFlow
+==================
+
 Install TensorFlow 1.14 from [yuhc/tensorflow-cudart-dynam](https://github.com/yuhc/tensorflow-cudart-dynam).
+Check out [BUILD.md](https://github.com/yuhc/tensorflow-cudart-dynam/blob/r1.14/BUILD.md) for build
+instructions.
+
+Build specification
+===================
+
+Configure AvA in `~/ava-build` with
+
+```shell
+cmake -DAVA_GEN_TF_SPEC=ON -DAVA_MANAGER_LEGACY=ON ../ava
+make
+```
+
+Run benchmark
+=============
+
+Dump CUDA binaries for the benchmark (Generally speaking, in most cases, CUDA
+binaries generated from a large TensorFlow benchmark can be used for most other
+benchmarks):
+
+```shell
+./install/bin/legacy_manager install/tf_dump/bin/worker
+LD_LIBRARY_PATH=~/ava-build/install/tf_dump/lib/ python3 your_tensorflow_benchmark.py
+sudo mkdir /cuda_dumps
+sudo cp /tmp/*.ava /cuda_dumps
+```
+
+> CUDA version: search `10.1` in `tf_dump.c` and `tf_opt.c`.
+> Dump path: search `cuda_dumps` in `tf_opt.c`.
+
+The CUDA dumps should be copied to both local (where benchmarks run) and remote
+(where API servers run) servers.
+
+Finally restart AvA manager and run the benchmark:
+
+```shell
+./install/bin/legacy_manager install/tf_opt/bin/worker
+LD_LIBRARY_PATH=~/ava-build/install/tf_opt/lib/ python3 your_tensorflow_benchmark.py
+```
diff --git a/cava/samples/tensorflow/tf_dump.c b/cava/samples/tensorflow/tf_dump.c
@@ -7565,8 +7565,13 @@ cudnnGetConvolutionBackwardDataWorkspaceSize(cudnnHandle_t handle,
                                              cudnnConvolutionBwdDataAlgo_t algo,
                                              size_t *sizeInBytes)
 {
-    fprintf(stderr, "%s is not implemented\n", __func__);
-    abort();
+   ava_argument(wDesc) ava_handle;
+   ava_argument(dyDesc) ava_handle;
+   ava_argument(convDesc) ava_handle;
+   ava_argument(dxDesc) ava_handle;
+   ava_argument(sizeInBytes) {
+     ava_out; ava_buffer(1);
+   }
 }
 
 cudnnStatus_t CUDNNWINAPI
@@ -7584,8 +7589,26 @@ cudnnConvolutionBackwardData(cudnnHandle_t handle,
                              const cudnnTensorDescriptor_t dxDesc,
                              void *dx)
 {
-    fprintf(stderr, "%s is not implemented\n", __func__);
-    abort();
+   ava_async;
+   ava_argument(handle) ava_handle;
+   ava_argument(alpha) {
+      ava_type_cast(const double *);
+      ava_in; ava_buffer(1);
+   }
+   ava_argument(beta) {
+      ava_type_cast(const double *);
+      ava_in; ava_buffer(1);
+   }
+   ava_argument(dx) ava_opaque;
+   ava_argument(dxDesc) ava_handle;
+   ava_argument(wDesc) ava_handle;
+   ava_argument(w) ava_opaque;
+   ava_argument(dyDesc) ava_handle;
+   ava_argument(dy) ava_opaque;
+   ava_argument(convDesc) ava_handle;
+   ava_argument(workSpace) {
+     ava_in; ava_buffer(workSpaceSizeInBytes);
+   }
 }
 
 cudnnStatus_t CUDNNWINAPI

diff --git a/cava/samples/tensorflow/tf_opt.c b/cava/samples/tensorflow/tf_opt.c
@@ -1168,7 +1168,7 @@ cudaLaunchKernel(const void *func, dim3 gridDim, dim3 blockDim, void **args,
         size_t sharedMem, cudaStream_t stream)
 {
     /* Cannot be ava_async, may lead to TensorFlow internal race condition */
-    ava_async;
+    // ava_async;
     ava_disable_native_call;
 
     ava_implicit_argument
@@ -8158,8 +8158,13 @@ cudnnGetConvolutionBackwardDataWorkspaceSize(cudnnHandle_t handle,
                                              cudnnConvolutionBwdDataAlgo_t algo,
                                              size_t *sizeInBytes)
 {
-    fprintf(stderr, "%s is not implemented\n", __func__);
-    abort();
+   ava_argument(wDesc) ava_handle;
+   ava_argument(dyDesc) ava_handle;
+   ava_argument(convDesc) ava_handle;
+   ava_argument(dxDesc) ava_handle;
+   ava_argument(sizeInBytes) {
+     ava_out; ava_buffer(1);
+   }
 }
 
 cudnnStatus_t CUDNNWINAPI
@@ -8177,8 +8182,26 @@ cudnnConvolutionBackwardData(cudnnHandle_t handle,
                              const cudnnTensorDescriptor_t dxDesc,
                              void *dx)
 {
-    fprintf(stderr, "%s is not implemented\n", __func__);
-    abort();
+   ava_async;
+   ava_argument(handle) ava_handle;
+   ava_argument(alpha) {
+      ava_type_cast(const double *);
+      ava_in; ava_buffer(1);
+   }
+   ava_argument(beta) {
+      ava_type_cast(const double *);
+      ava_in; ava_buffer(1);
+   }
+   ava_argument(dx) ava_opaque;
+   ava_argument(dxDesc) ava_handle;
+   ava_argument(wDesc) ava_handle;
+   ava_argument(w) ava_opaque;
+   ava_argument(dyDesc) ava_handle;
+   ava_argument(dy) ava_opaque;
+   ava_argument(convDesc) ava_handle;
+   ava_argument(workSpace) {
+     ava_in; ava_buffer(workSpaceSizeInBytes);
+   }
 }
 
 cudnnStatus_t CUDNNWINAPI

diff --git a/common/cmd_channel_socket_tcp.cpp b/common/cmd_channel_socket_tcp.cpp
@@ -24,7 +24,7 @@
 #include "common/cmd_handler.h"
 #include "cmd_channel_socket_utilities.h"
 #include "guest_config.h"
-#include "manager_service.pb.h"
+#include "manager_service.proto.h"
 
 using boost::asio::ip::tcp;
 
@@ -53,22 +53,25 @@ std::vector<struct command_channel*> command_channel_socket_tcp_guest_new()
 
     // Serialize configurations
     ava_proto::WorkerAssignRequest request;
-    request.set_gpu_count(guestconfig::config->gpu_memory_.size());
+    request.gpu_count() = guestconfig::config->gpu_memory_.size();
     for (auto m : guestconfig::config->gpu_memory_) {
-      request.add_gpu_mem(m << 20);
+      request.gpu_mem().push_back(m << 20);
     }
-    std::string request_buf(request.SerializeAsString());
-    uint32_t request_length = static_cast<uint32_t>(request_buf.length() + 1);
+    std::vector<unsigned char> request_buf;
+    zpp::serializer::memory_output_archive out(request_buf);
+    out(request);
+    uint32_t request_length = static_cast<uint32_t>(request_buf.size());
     boost::asio::write(manager_sock, boost::asio::buffer(&request_length, sizeof(uint32_t)));
-    boost::asio::write(manager_sock, boost::asio::buffer(request_buf.c_str(), request_length));
+    boost::asio::write(manager_sock, boost::asio::buffer(request_buf.data(), request_length));
 
     // De-serialize API server addresses
     uint32_t reply_length;
     boost::asio::read(manager_sock, boost::asio::buffer(&reply_length, sizeof(uint32_t)));
-    char reply_str[reply_length];
-    boost::asio::read(manager_sock, boost::asio::buffer(reply_str, reply_length));
+    std::vector<unsigned char> reply_buf(reply_length);
+    zpp::serializer::memory_input_archive in(reply_buf);
+    boost::asio::read(manager_sock, boost::asio::buffer(reply_buf.data(), reply_length));
     ava_proto::WorkerAssignReply reply;
-    reply.ParseFromString(reply_str);
+    in(reply);
     std::vector<std::string> worker_address;
     for (auto& wa : reply.worker_address()) {
         worker_address.push_back(wa);

diff --git a/docs/build_and_setup.md b/docs/build_and_setup.md
@@ -53,12 +53,13 @@ old codes into the new build system (v2.0) is still ongoing.
 
 > \* Upstream changes ([ava-serverless](https://github.com/photoszzt/ava-serverless)) have not been merged.
 
-| AvA manager | Status |
-| ----------- | ------ |
-| Demo        | TESTED |
-| Galvanic    | NO |
-| Katana      | NO |
-| Nvidia GPU  | NO |
+| AvA manager   | Status |
+| ------------- | ------ |
+| Demo          | NO |
+| Galvanic      | NO |
+| Katana        | NO |
+| Legacy (r1.0) | TESTED |
+| Nvidia GPU    | NO |
 
 This tutorial shows how to configure and build the minimal AvA demo API
 remoting system. To build and use other supported virtualized APIs, please