Feature: Support cuda_put()/get()

COPYRIGHT

@@ -27,20 +27,31 @@
 /*
  *  Ciprian Docan (2009)  TASSL Rutgers University
  *  [email protected]
+ *
  *  Tong Jin (2011) TASSL Rutgers University
  *  [email protected]
+ *
  *  Fan Zhang (2011) TASSL Rutgers University
  *  [email protected]
+ *
  *  Hoang Bui (2012) TASSL Rutgers University
  *  [email protected]
+ *
  *  Qian Sun (2012) TASSL Rutgers University
  *  [email protected]
+ *
  *  Melissa Romanus Abdelbaky (2014) TASSL Rutgers University
  *  [email protected]
+ *
  *  Mehmet Aktas
  *  [email protected]
+ *
  *  Pradeep Subedi (2017)  RDI2 Rutgers University
  *  [email protected]
+ *
  *  Philip Davis (2017) RDI2 Rutgers University
  *  [email protected]
+ *
+ *  Bo Zhang (2024) SCI Institute, University of Utah
+ *  [email protected]
  */

include/dspaces-common.h

@@ -27,6 +27,8 @@ extern "C" {
 #define dspaces_ERR_UNKNOWN_PR -7  /* Could not find server */
 #define dspaces_ERR_UNKNOWN_OBJ -8 /* Could not find the object*/
 #define dspaces_ERR_END -9         /* End of range for valid error codes */
+#define dspaces_ERR_CUDA -10         /* Error related to the CUDA */
+#define dspaces_ERR_UTILS -11
 
 #define DS_MOD_EFAULT -1
 #define DS_MOD_ENODEF -2

include/ss_data.h

@@ -22,6 +22,11 @@
 
 #include <abt.h>
 
+#ifdef HAVE_CUDA
+#include <cuda.h>
+#include <cuda_runtime_api.h>
+#endif /* HAVE_CUDA */
+
 #define MAX_VERSIONS 10
 
 #define DS_CLIENT_STORAGE 0x01
@@ -85,6 +90,26 @@ struct obj_data {
     unsigned int f_free : 1;
 };
 
+/*
+  A view in  the matrix allows to extract any subset  of values from a
+  matrix.
+*/
+
+struct matrix_view {
+    uint64_t lb[BBOX_MAX_NDIM];
+    uint64_t ub[BBOX_MAX_NDIM];
+};
+
+/* Generic matrix representation. */
+struct matrix {
+    uint64_t dist[BBOX_MAX_NDIM];
+    int num_dims;
+    size_t size_elem;
+    enum storage_type mat_storage;
+    struct matrix_view mat_view;
+    void *pdata;
+};
+
 struct gdim_list_entry {
     struct list_head entry;
     char *var_name;
@@ -355,7 +380,7 @@ static inline hg_return_t hg_proc_name_list_t(hg_proc_t proc, void *data)
         if(ret != HG_SUCCESS) {
             break;
         }
-        nlist->names = malloc(sizeof(*nlist->names) * nlist->count);
+        nlist->names = (hg_string_t*)malloc(sizeof(*nlist->names) * nlist->count);
         for(i = 0; i < nlist->count; i++) {
             ret = hg_proc_hg_string_t(proc, &nlist->names[i]);
             if(ret != HG_SUCCESS) {
@@ -411,7 +436,14 @@ int ssd_init(struct sspace *, int);
 void ssd_free(struct sspace *);
 //
 
+void matrix_init(struct matrix *, enum storage_type, struct bbox *, struct bbox *, void *, size_t);
+
 int ssd_copy(struct obj_data *, struct obj_data *);
+
+#ifdef HAVE_CUDA
+int ssd_copy_cuda(struct obj_data *, struct obj_data *);
+int ssd_copy_cuda_async(struct obj_data *, struct obj_data *, cudaStream_t *stream);
+#endif /* HAVE_CUDA*/
 //
 long ssh_hash_elem_count(struct sspace *ss, const struct bbox *bb);
 //
@@ -457,6 +489,11 @@ void obj_data_free(struct obj_data *od);
 uint64_t obj_data_size(obj_descriptor *);
 void obj_data_resize(obj_descriptor *obj_desc, uint64_t *new_dims);
 
+#ifdef HAVE_CUDA
+struct obj_data *obj_data_alloc_cuda(obj_descriptor *);
+void obj_data_free_cuda(struct obj_data *od);
+#endif /* HAVE_CUDA */
+
 int obj_desc_equals(obj_descriptor *, obj_descriptor *);
 int obj_desc_equals_no_owner(const obj_descriptor *, const obj_descriptor *);
 

include/util.h

@@ -2,11 +2,40 @@
 #define __DS_UTIL_H_
 
 #include <stdlib.h>
+#include <stdbool.h>
+#include <errno.h>
+#include "dspaces-common.h"
 
 size_t str_len(const char *str);
 char *str_append_const(char *, const char *);
 char *str_append(char *, char *);
 
+#define MEM_PATH_LEN 256
+typedef struct {
+    // Values from /proc/meminfo, in KiB or converted to MiB.
+    long long MemTotalKiB;
+    long long MemTotalMiB;
+    long long MemAvailableMiB; // -1 means no data available
+    long long SwapTotalMiB;
+    long long SwapTotalKiB;
+    long long SwapFreeMiB;
+    // Calculated percentages
+    double MemAvailablePercent; // percent of total memory that is available
+    double SwapFreePercent; // percent of total swap that is free
+} meminfo_t;
+
+typedef struct procinfo {
+    int pid;
+    int pidfd;
+    int uid;
+    int badness;
+    int oom_score_adj;
+    long long VmRSSkiB;
+    char name[MEM_PATH_LEN];
+} procinfo_t;
+
+meminfo_t parse_meminfo();
+
 /*******************************************************
    Processing parameter lists
 **********************************************************/

src/CMakeLists.txt

@@ -1,11 +1,35 @@
 find_package(DRC)
 if(HAVE_DRC)
     include_directories(${DRC_INCLUDE_DIRS})
-    add_definitions( -DHAVE_DRC)
+    add_definitions(-DHAVE_DRC)
+endif()
+
+# default: No CUDA
+set(HAVE_CUDA FALSE)
+if(CMAKE_VERSION VERSION_GREATER_EQUAL 3.17)
+    find_package(CUDAToolkit)
+    if(CUDAToolkit_FOUND)
+        enable_language(CUDA)
+        include_directories(${CUDAToolkit_INCLUDE_DIRS})
+        add_definitions(-DHAVE_CUDA)
+        set(HAVE_CUDA TRUE)
+    endif()
+else()
+    find_package(CUDA)
+    if(CUDA_FOUND)
+        enable_language(CUDA)
+        include_directories(${CUDA_INCLUDE_DIRS})
+        add_definitions(-DHAVE_CUDA)
+        set(HAVE_CUDA TRUE)
+    endif()
 endif()
 
 # list of source files
 set(dspaces-src util.c bbox.c ss_data.c dspaces-client.c dspaces-ops.c)
+if(HAVE_CUDA)
+    list(APPEND dspaces-src  ssd_copy.cu ss_data_cuda.c)
+    set_source_files_properties(ssd_copy.cu PROPERTIES LANGUAGE "CUDA")
+endif()
 
 # load package helper for generating cmake CONFIG packages
 include (CMakePackageConfigHelpers)
@@ -16,12 +40,30 @@ set (dspaces-pkg "share/cmake/dspaces")
 set (dspaces-vers "${dspaces_VERSION_MAJOR}.${dspaces_VERSION_MINOR}")
 
 add_library(dspaces ${dspaces-src})
+if(HAVE_CUDA)
+    set_target_properties(dspaces PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
+endif()
+
+# We need both cuda driver and cuda runtime headers and libs
+# ${CUDA_LIBRARIES} returned by find_package(CUDA) only has the cuda runtime libs
+# so we have to manually add -lcuda
+
+target_link_libraries (dspaces margo m pthread liblz4)
+target_include_directories (dspaces PUBLIC $<INSTALL_INTERFACE:include>)
+
 if(HAVE_DRC)
-    target_link_libraries (dspaces margo m pthread ${DRC_LIBRARIES} liblz4)
-    target_include_directories (dspaces PUBLIC $<INSTALL_INTERFACE:include> ${DRC_INCLUDE_DIRS})
-else()
-    target_link_libraries (dspaces margo m pthread liblz4)
-    target_include_directories (dspaces PUBLIC $<INSTALL_INTERFACE:include>)
+    target_link_libraries (dspaces ${DRC_LIBRARIES})
+    target_include_directories (dspaces PUBLIC ${DRC_INCLUDE_DIRS})
+endif()
+
+if(HAVE_CUDA)
+    if(CMAKE_VERSION VERSION_GREATER_EQUAL 3.17)
+        target_link_libraries (dspaces CUDA::cudart CUDA::cuda_driver)
+        target_include_directories (dspaces PUBLIC ${CUDAToolkit_INCLUDE_DIRS})
+    else()
+        target_link_libraries (dspaces cuda ${CUDA_LIBRARIES})
+        target_include_directories (dspaces PUBLIC ${CUDA_INCLUDE_DIRS})
+    endif()
 endif()
 
 # local include's BEFORE, in case old incompatable .h files in prefix/include
@@ -36,6 +78,11 @@ set_target_properties (dspaces
 set(dspaces-server-src util.c bbox.c ss_data.c dspaces-server.c dspaces-conf.c dspaces-modules.c dspaces-logging.c toml.c)
 
 add_library(dspaces-server ${dspaces-server-src} ${dspaces-src})
+
+target_link_libraries (dspaces-server margo m pthread liblz4 ${CURL_LIBRARIES})
+target_include_directories (dspaces-server BEFORE PUBLIC
+$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include> ${CURL_INCLUDE_DIR})
+
 if(HAVE_DRC)
     target_link_libraries (dspaces-server margo m pthread ${DRC_LIBRARIES} liblz4)
     target_include_directories (dspaces-server BEFORE PUBLIC