Implement the rdc_lib API to support the job stats

Add the function to start and stop the job recording.
Add the function to get the job stats for each GPU and summary of multiple GPUs
Add the function to remove the jobs.

Add a class RdcLogger which can control the log level using the environment variable RDC_LOG.
This is similar to GRPC_VERBOSITY gRPC. When the customer has the issues, he can enable the verbose
log to help us to troubleshoot the issues.

Add the -u support in the rdci group, fieldgroup and dmon for connecting to rdcd without authentication.

Change-Id: I22c591823c1ee6485db106b911bed8271d1b2769

rdc_libs/CMakeLists.txt

@@ -115,8 +115,10 @@ set(CMAKE_VERBOSE_MAKEFILE on)
 set(BOOTSTRAP_LIB "rdc_bootstrap")
 set(BOOTSTRAP_LIB_COMPONENT "lib${BOOTSTRAP_LIB}")
 set(BOOTSTRAP_LIB_SRC_LIST "${SRC_DIR}/bootstrap/src/RdcBootStrap.cc")
+set(BOOTSTRAP_LIB_SRC_LIST ${BOOTSTRAP_LIB_SRC_LIST} "${SRC_DIR}/bootstrap/src/RdcLogger.cc")
 set(BOOTSTRAP_LIB_INC_LIST "${RDC_LIB_INC_DIR}/rdc/rdc.h")
 set(BOOTSTRAP_LIB_INC_LIST ${BOOTSTRAP_LIB_INC_LIST} "${RDC_LIB_INC_DIR}/rdc_lib/rdc_common.h")
+set(BOOTSTRAP_LIB_INC_LIST ${BOOTSTRAP_LIB_INC_LIST} "${RDC_LIB_INC_DIR}/rdc_lib/RdcLogger.h")
 set(BOOTSTRAP_LIB_INC_LIST ${BOOTSTRAP_LIB_INC_LIST} "${RDC_LIB_INC_DIR}/rdc_lib/RdcHandler.h")
 
 message("BOOTSTRAP_LIB_INC_LIST=${BOOTSTRAP_LIB_INC_LIST}")

rdc_libs/bootstrap/src/RdcBootStrap.cc

@@ -24,6 +24,7 @@ THE SOFTWARE.
 #include <map>
 #include "rdc/rdc.h"
 #include "rdc_lib/RdcHandler.h"
+#include "rdc_lib/RdcLogger.h"
 #include "rdc_lib/rdc_common.h"
 
 static void* libHandler = nullptr;
@@ -96,7 +97,7 @@ rdc_status_t rdc_start_embedded(rdc_operation_mode_t op_mode,
 
         if (!libHandler) {
                 error = dlerror();
-                LOG_DEBUG("Fail to open librdc.so: " << error);
+                RDC_LOG(RDC_ERROR, "Fail to open librdc.so: " << error);
                 return RDC_ST_FAIL_LOAD_MODULE;
         }
 
@@ -104,7 +105,8 @@ rdc_status_t rdc_start_embedded(rdc_operation_mode_t op_mode,
                 dlsym(libHandler, "make_handler");
         if (!func_make_handler) {
                 error = dlerror();
-                LOG_DEBUG("Fail to find function make_handler:" << error);
+                RDC_LOG(RDC_ERROR,
+                "Fail to find function make_handler:" << error);
                 return RDC_ST_FAIL_LOAD_MODULE;
         }
 
@@ -144,15 +146,32 @@ rdc_status_t rdc_job_get_stats(rdc_handle_t p_rdc_handle, char  job_id[64] ,
 }
 
 rdc_status_t rdc_job_start_stats(rdc_handle_t p_rdc_handle,
-                rdc_gpu_group_t groupId, char  job_id[64], uint64_t update_freq,
-                double  max_keep_age, uint32_t  max_keep_samples ) {
+                rdc_gpu_group_t groupId, char  job_id[64],
+                 uint64_t update_freq) {
         if (!p_rdc_handle) {
                 return RDC_ST_INVALID_HANDLER;
         }
 
         return static_cast<amd::rdc::RdcHandler*>(p_rdc_handle)->
-                rdc_job_start_stats(groupId, job_id, update_freq,
-                                max_keep_age, max_keep_samples);
+                rdc_job_start_stats(groupId, job_id, update_freq);
+}
+
+rdc_status_t rdc_job_remove(rdc_handle_t p_rdc_handle, char job_id[64]) {
+        if (!p_rdc_handle) {
+                return RDC_ST_INVALID_HANDLER;
+        }
+
+        return static_cast<amd::rdc::RdcHandler*>(p_rdc_handle)->
+                rdc_job_remove(job_id);
+}
+
+rdc_status_t rdc_job_remove_all(rdc_handle_t p_rdc_handle) {
+        if (!p_rdc_handle) {
+                return RDC_ST_INVALID_HANDLER;
+        }
+
+        return static_cast<amd::rdc::RdcHandler*>(p_rdc_handle)->
+                rdc_job_remove_all();
 }
 
 
@@ -344,6 +363,8 @@ const char* rdc_status_string(rdc_status_t result) {
                 return "The max limit reached";
         case RDC_ST_CONFLICT:
                 return "Conflict with current state";
+        case RDC_ST_ALREADY_EXIST:
+                return "The value already exists";
         case RDC_ST_CLIENT_ERROR:
                 return "RDC Client error";
         default:

rdc_libs/bootstrap/src/RdcLogger.cc

@@ -0,0 +1,78 @@
+/*
+Copyright (c) 2020 - present Advanced Micro Devices, Inc. All rights reserved.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in
+all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+THE SOFTWARE.
+*/
+#include "rdc_lib/RdcLogger.h"
+#include <stdlib.h>
+#include <string.h>
+#include <sstream>
+#include <iomanip>
+#include <iostream>
+#include <chrono>  // NOLINT
+#include "rdc_lib/rdc_common.h"
+
+namespace amd {
+namespace rdc {
+
+RdcLogger::RdcLogger(std::ostream& os):
+    os_(os) {
+    char* verbose = getenv("RDC_LOG");
+    if (verbose == nullptr) {
+        log_level_ = RDC_ERROR;
+    } else if (strcmp(verbose, "DEBUG") == 0) {
+       log_level_ = RDC_DEBUG;
+    } else if (strcmp(verbose, "INFO") == 0) {
+       log_level_ = RDC_INFO;
+    } else {
+       log_level_ = RDC_ERROR;
+    }
+}
+
+std::string RdcLogger::get_log_header(uint32_t severity,
+            const char* file, int line) {
+    std::stringstream strstream;
+    auto ms =  std::chrono::duration_cast<std::chrono::milliseconds>
+         (std::chrono::system_clock::now().time_since_epoch()).count();
+    strstream << std::fixed << std::setprecision(3) << (ms/1000.0) << " ";
+    if (severity == RDC_DEBUG) {
+       strstream << "DEBUG ";
+    } else if (severity == RDC_INFO) {
+       strstream << "INFO ";
+    } else {
+       strstream << "ERROR ";
+    }
+
+    //  extract out the file path as it may be very long.
+    if (file != nullptr) {
+       std::string file_str(file);
+       auto found = file_str.find_last_of("/");
+       if (found != std::string::npos) {
+          file_str = file_str.substr(found+1);
+       }
+       strstream << file_str << "(" << line << "): ";
+    }
+
+    return strstream.str();
+}
+
+
+}  // namespace rdc
+}  // namespace amd
+

rdc_libs/rdc/src/RdcCacheManagerImpl.cc

@@ -22,6 +22,8 @@ THE SOFTWARE.
 #include "rdc_lib/impl/RdcCacheManagerImpl.h"
 #include <sys/time.h>
 #include <ctime>
+#include <sstream>
+#include "rdc_lib/RdcLogger.h"
 #include "rdc_lib/rdc_common.h"
 
 
@@ -127,15 +129,26 @@ rdc_status_t RdcCacheManagerImpl::rdc_field_get_latest_value(
     return RDC_ST_OK;
 }
 
-uint32_t RdcCacheManagerImpl::get_cache_size() {
-    uint32_t cache_size = 0;
+std::string RdcCacheManagerImpl::get_cache_stats() {
+    std::stringstream strstream;
     std::lock_guard<std::mutex> guard(cache_mutex_);
 
+    strstream << "Cache samples:";
     auto cache_samples_ite = cache_samples_.begin();
     for (; cache_samples_ite != cache_samples_.end(); cache_samples_ite++) {
-        cache_size+=cache_samples_ite->second.size();
+        strstream << "<" << cache_samples_ite->first.first << ","
+            << cache_samples_ite->first.second << ":"
+            << cache_samples_ite->second.size() << "> ";
     }
-    return cache_size;
+
+    strstream <<" Job caches:";
+    auto job_ite = cache_jobs_.begin();
+    for ( ; job_ite != cache_jobs_.end(); job_ite++ ) {
+        strstream << "<" << job_ite->first << ":"
+            << job_ite->second.gpu_stats.size() << "> ";
+    }
+
+    return strstream.str();
 }
 
 rdc_status_t RdcCacheManagerImpl::rdc_update_cache(uint32_t gpu_index,
@@ -162,5 +175,188 @@ rdc_status_t RdcCacheManagerImpl::rdc_update_cache(uint32_t gpu_index,
     return RDC_ST_OK;
 }
 
+rdc_status_t RdcCacheManagerImpl::rdc_job_remove(char job_id[64]) {
+    std::lock_guard<std::mutex> guard(cache_mutex_);
+    cache_jobs_.erase(job_id);
+    return RDC_ST_OK;
+}
+
+rdc_status_t RdcCacheManagerImpl::rdc_job_remove_all() {
+    std::lock_guard<std::mutex> guard(cache_mutex_);
+    cache_jobs_.clear();
+    return RDC_ST_OK;
+}
+
+rdc_status_t RdcCacheManagerImpl::rdc_update_job_stats(uint32_t gpu_index,
+    const std::string& job_id, const rdc_field_value& value) {
+    std::lock_guard<std::mutex> guard(cache_mutex_);
+    auto job_iter = cache_jobs_.find(job_id);
+    if (job_iter == cache_jobs_.end()) {
+        return RDC_ST_NOT_FOUND;
+    }
+
+    auto gpu_iter = job_iter->second.gpu_stats.find(gpu_index);
+    if (gpu_iter == job_iter->second.gpu_stats.end()) {
+        return RDC_ST_NOT_FOUND;
+    }
+
+    auto fsummary = gpu_iter->second.field_summaries.find(value.field_id);
+    if (fsummary == gpu_iter->second.field_summaries.end()) {
+        return RDC_ST_NOT_FOUND;
+    }
+    if (fsummary->second.count == 0) {  // first item
+        fsummary->second.count = 1;
+        fsummary->second.max_value = value.value.l_int;
+        fsummary->second.min_value =  value.value.l_int;
+        fsummary->second.total_value = value.value.l_int;
+        fsummary->second.last_time = value.ts;
+        if (value.field_id == RDC_FI_POWER_USAGE) {
+            gpu_iter->second.energy_last_time = value.ts;
+        }
+        return RDC_ST_OK;
+    }
+    if (value.field_id == RDC_FI_POWER_USAGE) {
+        uint64_t time_elapsed = value.ts - gpu_iter->second.energy_last_time;
+        // Stored in cache as microseconds and microwats
+        gpu_iter->second.energy_consumed +=
+            (time_elapsed * value.value.l_int)/(1000.0*1000000);
+    }
+    fsummary->second.max_value = std::max(fsummary->second.max_value,
+        static_cast<int64_t>(value.value.l_int));
+    fsummary->second.min_value = std::min(fsummary->second.min_value,
+        static_cast<int64_t>(value.value.l_int));
+    fsummary->second.total_value += value.value.l_int;
+    fsummary->second.last_time = value.ts;
+    fsummary->second.count++;
+
+    return RDC_ST_OK;
+}
+
+void RdcCacheManagerImpl::set_summary(const FieldSummaryStats & stats,
+    rdc_stats_summary_t & gpu, rdc_stats_summary_t& summary,
+    unsigned int adjuster) {
+    if (stats.count == 0) return;
+
+    gpu.max_value = stats.max_value / adjuster;
+    gpu.min_value = stats.min_value / adjuster;
+    gpu.average = stats.total_value / stats.count / adjuster;
+    summary.max_value = std::max(summary.max_value, gpu.max_value);
+    summary.min_value = std::min(summary.min_value, gpu.min_value);
+    //< save total for future average calculation.
+    summary.average += gpu.average;
+}
+
+rdc_status_t RdcCacheManagerImpl::rdc_job_get_stats(char jobId[64],
+        const rdc_gpu_total_memory_t& total_memory,
+        rdc_job_info_t* p_job_info) {
+    std::lock_guard<std::mutex> guard(cache_mutex_);
+    auto job_stats = cache_jobs_.find(jobId);
+
+    if (job_stats == cache_jobs_.end()) {
+        return RDC_ST_NOT_FOUND;
+    }
+
+    //< Init the summary info
+    RDC_LOG(RDC_DEBUG, "rdc_job_get_stats for job "  <<jobId);
+    auto& summary_info = p_job_info->summary;
+    summary_info.start_time = job_stats->second.start_time;
+    if (job_stats->second.end_time == 0) {
+        summary_info.end_time = time(nullptr);
+    } else {
+        summary_info.end_time = job_stats->second.end_time;
+    }
+    summary_info.energy_consumed = 0;
+    summary_info.max_gpu_memory_used = 0;
+    summary_info.power_usage = {0, std::numeric_limits<uint64_t>::max(), 0};
+    summary_info.gpu_clock = {0, std::numeric_limits<uint64_t>::max(), 0};
+    summary_info.gpu_utilization = {0, std::numeric_limits<uint64_t>::max(), 0};
+    summary_info.memory_utilization = {0,
+                    std::numeric_limits<uint64_t>::max(), 0};
+
+    p_job_info->num_gpus = job_stats->second.gpu_stats.size();
+
+    //< Populate information for each GPUs
+
+    auto gpus = job_stats->second.gpu_stats.begin();
+    for (; gpus != job_stats->second.gpu_stats.end(); gpus++) {
+        auto & gpu_info = p_job_info->gpus[gpus->first];
+        gpu_info.start_time = summary_info.start_time;
+        gpu_info.end_time = summary_info.end_time;
+        gpu_info.energy_consumed = gpus->second.energy_consumed;
+        summary_info.energy_consumed += gpu_info.energy_consumed;
+
+        auto ite = gpus->second.field_summaries.begin();
+        for (; ite != gpus->second.field_summaries.end(); ite++) {
+            if (ite->first == RDC_FI_POWER_USAGE) {
+                set_summary(ite->second,
+                    gpu_info.power_usage, summary_info.power_usage, 1000000);
+            } else if (ite->first == RDC_FI_GPU_MEMORY_USAGE) {
+                auto tmemory = total_memory.at(gpus->first);
+                set_summary(ite->second, gpu_info.memory_utilization,
+                    summary_info.memory_utilization, tmemory/100);
+                gpu_info.max_gpu_memory_used = ite->second.max_value;
+                summary_info.max_gpu_memory_used = std::max(
+                    summary_info.max_gpu_memory_used,
+                    gpu_info.max_gpu_memory_used);
+            } else if (ite->first == RDC_FI_GPU_SM_CLOCK) {
+                set_summary(ite->second, gpu_info.gpu_clock,
+                    summary_info.gpu_clock, 1000000);
+            } else if (ite->first == RDC_FI_GPU_UTIL) {
+                set_summary(ite->second, gpu_info.gpu_utilization,
+                    summary_info.gpu_utilization, 1);
+            }
+        }
+    }
+    // Get the average of the summary
+    summary_info.power_usage.average = summary_info.power_usage.average/
+                p_job_info->num_gpus;
+    summary_info.gpu_clock.average = summary_info.gpu_clock.average/
+                p_job_info->num_gpus;
+    summary_info.gpu_utilization.average = summary_info.gpu_utilization.average/
+                p_job_info->num_gpus;
+    summary_info.memory_utilization.average =
+                summary_info.memory_utilization.average/p_job_info->num_gpus;
+
+    return RDC_ST_OK;
+}
+
+rdc_status_t RdcCacheManagerImpl::rdc_job_start_stats(char job_id[64],
+        const rdc_group_info_t& ginfo, const rdc_field_group_info_t& finfo) {
+     RdcJobStatsCacheEntry cacheEntry;
+     cacheEntry.start_time = std::time(nullptr);
+     cacheEntry.end_time = 0;
+     for (unsigned int i=0 ; i < ginfo.count; i++) {  // GPUs
+       GpuSummaryStats gstats;
+       gstats.energy_consumed = 0;
+       gstats.energy_last_time = 0;
+       for (unsigned int j = 0; j < finfo.count; j++) {  // init fields
+          FieldSummaryStats s;
+          s.count = 0;
+          s.max_value = s.min_value = s.total_value = 0;
+          gstats.field_summaries.insert({finfo.field_ids[j], s});
+       }
+
+       cacheEntry.gpu_stats.insert({ginfo.entity_ids[i], gstats});
+     }
+
+     std::lock_guard<std::mutex> guard(cache_mutex_);
+     cache_jobs_.insert({job_id, cacheEntry});
+     return RDC_ST_OK;
+}
+
+
+rdc_status_t RdcCacheManagerImpl::rdc_job_stop_stats(char job_id[64]) {
+    std::lock_guard<std::mutex> guard(cache_mutex_);
+    auto job_stats = cache_jobs_.find(job_id);
+
+    if (job_stats == cache_jobs_.end()) {
+        return RDC_ST_NOT_FOUND;
+    }
+
+    job_stats->second.end_time = std::time(nullptr);
+
+    return RDC_ST_OK;
+}
+
 }  // namespace rdc
 }  // namespace amd

rdc_libs/rdc/src/RdcEmbeddedHandler.cc

@@ -27,6 +27,7 @@ THE SOFTWARE.
 #include "rdc_lib/impl/RdcCacheManagerImpl.h"
 #include "rdc_lib/impl/RdcWatchTableImpl.h"
 #include "rdc_lib/rdc_common.h"
+#include "rdc_lib/RdcLogger.h"
 #include "rdc_lib/RdcException.h"
 #include "rocm_smi/rocm_smi.h"
 
@@ -73,6 +74,7 @@ RdcEmbeddedHandler::RdcEmbeddedHandler(rdc_operation_mode_t mode):
     , metrics_updater_(new RdcMetricsUpdaterImpl(watch_table_,
                         METIC_UPDATE_FREQUENCY)) {
     if (mode == RDC_OPERATION_MODE_AUTO) {
+        RDC_LOG(RDC_DEBUG, "Run RDC with RDC_OPERATION_MODE_AUTO");
         metrics_updater_->start();
     }
 }
@@ -83,32 +85,49 @@ RdcEmbeddedHandler::~RdcEmbeddedHandler() {
 
 // JOB API
 rdc_status_t RdcEmbeddedHandler::rdc_job_start_stats(rdc_gpu_group_t groupId,
-        char job_id[64], uint64_t update_freq, double  max_keep_age,
-        uint32_t  max_keep_samples) {
-    // TODO(bill_liu): implement
-    (void)(groupId);
-    (void)(job_id);
-    (void)(update_freq);
-    (void)(max_keep_age);
-    (void)(max_keep_samples);
-
-    return RDC_ST_OK;
+        char job_id[64], uint64_t update_freq) {
+    return  watch_table_->rdc_job_start_stats(groupId, job_id, update_freq);
 }
 
-rdc_status_t RdcEmbeddedHandler::rdc_job_get_stats(char  job_id[64],
+rdc_status_t RdcEmbeddedHandler::rdc_job_get_stats(char job_id[64],
            rdc_job_info_t* p_job_info) {
-    // TODO(bill_liu): implement
-    (void)(job_id);
-    (void)(p_job_info);
-    return RDC_ST_OK;
+    uint32_t gpu_index_list[RDC_MAX_NUM_DEVICES];
+    uint32_t count = 0;
+    rdc_status_t status = rdc_device_get_all(
+        gpu_index_list, &count);
+    if (status != RDC_ST_OK) {
+        return status;
+    }
+
+    rdc_gpu_total_memory_t all_total_memory;
+
+    for (uint32_t i = 0; i < count ; i++) {
+        rdc_field_value total_memory;
+        status = metric_fetcher_->fetch_smi_field(gpu_index_list[i],
+                    RDC_FI_GPU_MEMORY_TOTAL, &total_memory);
+        if (status != RDC_ST_OK) {
+            RDC_LOG(RDC_ERROR, "Fail to get total memory of GPU "
+                        << gpu_index_list[i]);
+            return status;
+        }
+        all_total_memory.insert({gpu_index_list[i], total_memory.value.l_int});
+    }
+
+    return cache_mgr_->rdc_job_get_stats(job_id, all_total_memory, p_job_info);
 }
 
-rdc_status_t RdcEmbeddedHandler::rdc_job_stop_stats(char  job_id[64] ) {
-    // TODO(bill_liu): implement
-    (void)(job_id);
-    return RDC_ST_OK;
+rdc_status_t RdcEmbeddedHandler::rdc_job_stop_stats(char job_id[64]) {
+    return watch_table_->rdc_job_stop_stats(job_id);
 }
 
+rdc_status_t RdcEmbeddedHandler::rdc_job_remove(char job_id[64]) {
+    return watch_table_->rdc_job_remove(job_id);
+}
+
+
+rdc_status_t RdcEmbeddedHandler::rdc_job_remove_all() {
+    return watch_table_->rdc_job_remove_all();
+}
 
 // Discovery API
 rdc_status_t RdcEmbeddedHandler::rdc_device_get_all(
@@ -194,6 +213,8 @@ rdc_status_t RdcEmbeddedHandler::rdc_group_gpu_add(rdc_gpu_group_t group_id,
     }
 
     if (!is_gpu_exist) {
+        RDC_LOG(RDC_INFO, "Fail to add GPU index " << gpu_index << " to group "
+            << group_id <<" as the GPU index is invalid.");
         return RDC_ST_NOT_FOUND;
     }
 
@@ -211,6 +232,9 @@ rdc_status_t RdcEmbeddedHandler::rdc_group_field_create(uint32_t num_field_ids,
     if (num_field_ids <= RDC_MAX_FIELD_IDS_PER_FIELD_GROUP) {
         for (uint32_t i = 0; i < num_field_ids; i++) {
             if (!metric_fetcher_->is_field_valid(field_ids[i])) {
+                RDC_LOG(RDC_INFO,
+                    "Fail to create field group with unknown field id "
+                    << field_ids[i]);
                 return RDC_ST_NOT_SUPPORTED;
             }
         }
@@ -285,6 +309,9 @@ rdc_status_t RdcEmbeddedHandler::rdc_field_get_latest_value(
         return RDC_ST_BAD_PARAMETER;
     }
     if (!metric_fetcher_->is_field_valid(field)) {
+        RDC_LOG(RDC_INFO,
+                    "Fail to get latest value with unknown field id "
+                    << field);
         return RDC_ST_NOT_SUPPORTED;
     }
     return cache_mgr_->rdc_field_get_latest_value(gpu_index, field, value);
@@ -297,6 +324,9 @@ rdc_status_t RdcEmbeddedHandler::rdc_field_get_value_since(uint32_t gpu_index,
         return RDC_ST_BAD_PARAMETER;
     }
     if (!metric_fetcher_->is_field_valid(field)) {
+        RDC_LOG(RDC_INFO,
+                "Fail to get value since with unknown field id "
+                << field);
         return RDC_ST_NOT_SUPPORTED;
     }
     return cache_mgr_->rdc_field_get_value_since(gpu_index, field,

rdc_libs/rdc/src/RdcGroupSettingsImpl.cc

@@ -22,11 +22,20 @@ THE SOFTWARE.
 #include "rdc_lib/impl/RdcGroupSettingsImpl.h"
 #include <ctime>
 #include "rdc_lib/rdc_common.h"
+#include "rdc_lib/RdcLogger.h"
 
 namespace amd {
 namespace rdc {
 
 RdcGroupSettingsImpl::RdcGroupSettingsImpl() {
+    // Add the default job stats fields
+    uint32_t job_fields[] = {RDC_FI_GPU_MEMORY_USAGE,
+        RDC_FI_POWER_USAGE, RDC_FI_GPU_SM_CLOCK, RDC_FI_GPU_UTIL};
+    char job_field_group[] = "JobStatsFields";
+    rdc_field_grp_t fgid = JOB_FIELD_ID;
+
+    rdc_group_field_create(sizeof(job_fields)/sizeof(uint32_t),
+                job_fields,  job_field_group, &fgid);
 }
 
 rdc_status_t RdcGroupSettingsImpl::rdc_group_gpu_create(
@@ -62,6 +71,8 @@ rdc_status_t RdcGroupSettingsImpl::rdc_group_gpu_add(
         // Check whether the index already exists
         for (uint32_t i=0; i < ite->second.count; i++) {
             if (ite->second.entity_ids[i] == gpu_index) {
+                RDC_LOG(RDC_INFO, "Fail to add " << gpu_index
+                <<" to GPU group " << groupId << " as it is already exists");
                 return RDC_ST_BAD_PARAMETER;
             }
         }
@@ -136,15 +147,19 @@ rdc_status_t RdcGroupSettingsImpl::rdc_group_field_create(
     if (field_group_.size() >= RDC_MAX_NUM_FIELD_GROUPS) {
         return RDC_ST_MAX_LIMIT;
     }
-    field_group_.emplace(cur_filed_group_id_, finfo);
-    *rdc_field_group_id = cur_filed_group_id_;
-    cur_filed_group_id_++;
+    field_group_.emplace(cur_field_group_id_, finfo);
+    *rdc_field_group_id = cur_field_group_id_;
+    cur_field_group_id_++;
 
     return RDC_ST_OK;
 }
 
 rdc_status_t RdcGroupSettingsImpl::rdc_group_field_destroy(
     rdc_field_grp_t rdc_field_group_id) {
+    if (rdc_field_group_id == JOB_FIELD_ID) {
+        RDC_LOG(RDC_INFO, "Cannot delete system JOB_FIELD_ID field group");
+        return RDC_ST_BAD_PARAMETER;
+    }
     std::lock_guard<std::mutex> guard(field_group_mutex_);
     field_group_.erase(rdc_field_group_id);
     return RDC_ST_OK;
@@ -183,6 +198,10 @@ rdc_status_t RdcGroupSettingsImpl::rdc_group_field_get_all_ids(
         if (*count >= RDC_MAX_NUM_FIELD_GROUPS) {
             return RDC_ST_MAX_LIMIT;
         }
+
+        // Skip system defined JOB_FIELD_ID
+        if (ite->first == JOB_FIELD_ID) continue;
+
         field_group_id_list[*count] = ite->first;
         (*count)++;
     }

rdc_libs/rdc/src/RdcMetricFetcherImpl.cc

@@ -22,10 +22,11 @@ THE SOFTWARE.
 #include "rdc_lib/impl/RdcMetricFetcherImpl.h"
 #include <sys/time.h>
 #include <string.h>
-#include <chrono>
+#include <chrono>  //NOLINT
 #include <algorithm>
 #include <vector>
 #include "rdc_lib/rdc_common.h"
+#include "rdc_lib/RdcLogger.h"
 #include "rocm_smi/rocm_smi.h"
 
 namespace amd {
@@ -48,6 +49,8 @@ rdc_status_t RdcMetricFetcherImpl::fetch_smi_field(uint32_t gpu_index,
     uint64_t i64 = 0;
 
     if (!is_field_valid(field_id)) {
+         RDC_LOG(RDC_ERROR, "Fail to fetch field " << field_id
+              << " which is not supported");
          return RDC_ST_NOT_SUPPORTED;
     }
 
@@ -125,6 +128,27 @@ rdc_status_t RdcMetricFetcherImpl::fetch_smi_field(uint32_t gpu_index,
             break;
     }
 
+    gettimeofday(&tv, NULL);
+    int64_t latency = static_cast<uint64_t>(tv.tv_sec)*1000+tv.tv_usec/1000
+               - value->ts;
+    if (value->status != RSMI_STATUS_SUCCESS) {
+         RDC_LOG(RDC_ERROR, "Fail to fetch " << gpu_index << ":" <<
+              field_id_string(field_id) << " with rsmi error code "
+              << value->status <<", latency " << latency);
+    } else if (value->type == INTEGER) {
+         RDC_LOG(RDC_DEBUG, "Fetch " << gpu_index << ":" <<
+               field_id_string(field_id) << ":" << value->value.l_int
+               << ", latency " << latency);
+    } else if (value->type == DOUBLE) {
+         RDC_LOG(RDC_DEBUG, "Fetch " << gpu_index << ":" <<
+         field_id_string(field_id) << ":" << value->value.dbl
+         << ", latency " << latency);
+    } else if (value->type == STRING) {
+         RDC_LOG(RDC_DEBUG, "Fetch " << gpu_index << ":" <<
+         field_id_string(field_id) << ":" << value->value.str
+         << ", latency " << latency);
+    }
+
     return value->status == RSMI_STATUS_SUCCESS ?  RDC_ST_OK : RDC_ST_MSI_ERROR;
 }
 

rdc_libs/rdc/src/RdcWatchTableImpl.cc

@@ -23,8 +23,10 @@ THE SOFTWARE.
 #include "rdc_lib/impl/RdcWatchTableImpl.h"
 #include <sys/time.h>
 #include <ctime>
+#include <sstream>
 #include <algorithm>
 #include "rdc_lib/rdc_common.h"
+#include "rdc_lib/RdcLogger.h"
 
 namespace amd {
 namespace rdc {
@@ -39,24 +41,92 @@ RdcWatchTableImpl::RdcWatchTableImpl(const RdcGroupSettingsPtr& group_settings,
 }
 
 rdc_status_t  RdcWatchTableImpl::rdc_job_start_stats(rdc_gpu_group_t group_id,
-                char  job_id[64]) {
-    // TODO(bill_liu): implement
-    (void)(group_id);
-    (void)(job_id);
-    return RDC_ST_OK;
+                char  job_id[64], uint64_t update_freq) {
+    do {  //< lock guard for thread safe
+        std::lock_guard<std::mutex> guard(watch_mutex_);
+        if (job_watch_table_.find(job_id) != job_watch_table_.end()) {
+            return RDC_ST_ALREADY_EXIST;
+        }
+    } while (0);
+
+    std::vector<RdcFieldKey> fields_in_watch;
+    rdc_status_t result = get_fields_from_group(group_id,
+                JOB_FIELD_ID, fields_in_watch);
+    JobWatchTableEntry jentry {group_id, fields_in_watch};
+    job_watch_table_.insert({job_id, jentry});
+
+    result = rdc_field_watch(group_id, JOB_FIELD_ID, update_freq, 0, 0);
+    if (result != RDC_ST_OK) {
+        return result;
+    }
+
+    rdc_field_group_info_t finfo;
+    rdc_group_info_t ginfo;
+    result = group_settings_->rdc_group_gpu_get_info(group_id, &ginfo);
+    if (result != RDC_ST_OK) {
+        return result;
+    }
+    result = group_settings_->rdc_group_field_get_info(JOB_FIELD_ID, &finfo);
+    if (result != RDC_ST_OK) {
+        return result;
+    }
+
+    result = cache_mgr_->rdc_job_start_stats(job_id, ginfo, finfo);
+
+    return result;
 }
 
-rdc_status_t RdcWatchTableImpl::rdc_watch_job_fields(rdc_gpu_group_t group_id,
-                uint64_t update_freq, double  max_keep_age,
-                uint32_t max_keep_samples) {
-    // TODO(bill_liu): implement
-    (void)(group_id);
-    (void)(update_freq);
-    (void)(max_keep_age);
-    (void)(max_keep_samples);
-    return RDC_ST_OK;
+rdc_status_t RdcWatchTableImpl::rdc_job_stop_stats(char job_id[64]) {
+    uint32_t job_group_id;
+    do {  //< lock guard for thread safe
+        std::lock_guard<std::mutex> guard(watch_mutex_);
+        auto job = job_watch_table_.find(job_id);
+        if (job == job_watch_table_.end()) {
+            return RDC_ST_NOT_FOUND;
+        }
+        job_group_id = job->second.group_id;
+    } while (0);
+
+    rdc_status_t result = rdc_field_unwatch(job_group_id, JOB_FIELD_ID);
+    if (result != RDC_ST_OK) {
+        return result;
+    }
+
+    do {  //< lock guard for thread safe
+        std::lock_guard<std::mutex> guard(watch_mutex_);
+        job_watch_table_.erase(job_id);
+    } while (0);
+
+    result = cache_mgr_->rdc_job_stop_stats(job_id);
+
+    return result;
 }
 
+rdc_status_t RdcWatchTableImpl::rdc_job_remove(char job_id[64]) {
+    rdc_job_stop_stats(job_id);
+    return cache_mgr_->rdc_job_remove(job_id);
+}
+
+rdc_status_t RdcWatchTableImpl::rdc_job_remove_all() {
+    // Get all the job ids;
+    std::vector<std::string> v;
+    do {  //< lock guard for thread safe
+        std::lock_guard<std::mutex> guard(watch_mutex_);
+        for (auto ite = job_watch_table_.begin();
+            ite != job_watch_table_.end(); ite++) {
+            v.push_back(ite->first);
+        }
+    } while (0);
+
+    // Stop them
+    for (auto job = v.begin(); job != v.end(); job++) {
+        rdc_job_stop_stats(const_cast<char*>(job->c_str()));
+    }
+
+    return cache_mgr_->rdc_job_remove_all();
+}
+
+
 rdc_status_t RdcWatchTableImpl::get_fields_from_group(rdc_gpu_group_t group_id,
     rdc_field_grp_t field_group_id, std::vector<RdcFieldKey> & fields) {
     rdc_field_group_info_t finfo;
@@ -228,6 +298,21 @@ rdc_status_t RdcWatchTableImpl::rdc_field_unwatch(
     return update_field_in_table_when_unwatch(ite->first);
 }
 
+bool RdcWatchTableImpl::is_job_watch_field(uint32_t gpu_index,
+        uint32_t field_id, std::string& job_id) const {
+    RdcFieldKey key{gpu_index, field_id};
+
+    for (auto ite = job_watch_table_.begin();
+                ite != job_watch_table_.end(); ite++) {
+        auto& fields = ite->second.fields;
+        if (std::find(fields.begin(), fields.end(), key) != fields.end()) {
+            job_id = ite->first;
+            return true;
+        }
+    }
+
+    return false;
+}
 
 rdc_status_t RdcWatchTableImpl::rdc_field_update_all() {
     uint32_t items_fetched = 0;
@@ -251,13 +336,19 @@ rdc_status_t RdcWatchTableImpl::rdc_field_update_all() {
         result = metric_fetcher_->fetch_smi_field(
                     fite->first.first, fite->first.second, &value);
         if (result != RDC_ST_OK) {
-           LOG_DEBUG("Fail to fetch the field: " << rdc_status_string(result));
            continue;
         }
 
         // Update the cache
         cache_mgr_->rdc_update_cache(fite->first.first, value);
 
+        // Update the job stats cache
+        std::string job_id;
+        if (is_job_watch_field(fite->first.first, fite->first.second, job_id)) {
+            cache_mgr_->rdc_update_job_stats(fite->first.first, job_id, value);
+        }
+
+
         // Update the last_upate_time
         gettimeofday(&tv, NULL);
         now = static_cast<uint64_t>(tv.tv_sec)*1000+tv.tv_usec/1000;
@@ -303,6 +394,56 @@ void RdcWatchTableImpl::clean_up() {
             ++wite;
         }
     }
+
+    // Debug log every 30 seconds
+    if (now/1000%30 == 0) {
+        debug_status();
+    }
+}
+
+void RdcWatchTableImpl::debug_status() {
+    RDC_LOG(RDC_DEBUG, "fields_to_watch_:" << fields_to_watch_.size()
+            << " watch_table_:" << watch_table_.size()
+            << " job_watch_table_:" << job_watch_table_.size()
+            << " cache stats:" << cache_mgr_->get_cache_stats());
+
+    if (watch_table_.size() > 0) {
+        RDC_LOG(RDC_DEBUG, "watch table details:");
+    }
+    for (auto wite = watch_table_.begin(); wite != watch_table_.end(); wite++) {
+        RDC_LOG(RDC_DEBUG, wite->first.first << "," << wite->first.second
+            << ": age:" << wite->second.max_keep_age << ", samples:"
+            << wite->second.max_keep_samples << ", is_watching:"
+            << wite->second.is_watching << ", last_update_time:"
+            << wite->second.last_update_time <<", update_freq:"
+            << wite->second.update_freq);
+    }
+
+    if (job_watch_table_.size() > 0) {
+        RDC_LOG(RDC_DEBUG, "job watch table details: ");
+    }
+    for (auto jite = job_watch_table_.begin();
+            jite !=job_watch_table_.end(); jite++) {
+        std::stringstream strstream;
+        for (const auto& p : jite->second.fields) {
+            strstream << "<" << p.first << "," << p.second << "> ";
+        }
+        RDC_LOG(RDC_DEBUG, jite->first << ": " << jite->second.group_id
+                << " fields : "<< strstream.str());
+    }
+
+    if (fields_to_watch_.size() > 0) {
+        RDC_LOG(RDC_DEBUG, "fields to watch details:");
+    }
+    for (auto fite = fields_to_watch_.begin(); fite != fields_to_watch_.end();
+            fite++) {
+        RDC_LOG(RDC_DEBUG, fite->first.first << "," << fite->first.second
+            << ": age:" << fite->second.max_keep_age << ", samples:"
+            << fite->second.max_keep_samples << ", is_watching:"
+            << fite->second.is_watching << ", last_update_time:"
+            << fite->second.last_update_time <<", update_freq:"
+            << fite->second.update_freq);
+    }
 }
 
 }  // namespace rdc

rdc_libs/rdc_client/src/RdcStandaloneHandler.cc

@@ -62,19 +62,16 @@ rdc_status_t RdcStandaloneHandler::error_handle(::grpc::Status status,
 
 // JOB RdcAPI
 rdc_status_t RdcStandaloneHandler::rdc_job_start_stats(rdc_gpu_group_t groupId,
-        char job_id[64], uint64_t update_freq, double  max_keep_age,
-        uint32_t  max_keep_samples) {
+        char job_id[64], uint64_t update_freq) {
     // TODO(bill_liu): implement
     (void)(groupId);
     (void)(job_id);
     (void)(update_freq);
-    (void)(max_keep_age);
-    (void)(max_keep_samples);
 
     return RDC_ST_OK;
 }
 
-rdc_status_t RdcStandaloneHandler::rdc_job_get_stats(char  job_id[64],
+rdc_status_t RdcStandaloneHandler::rdc_job_get_stats(char job_id[64],
            rdc_job_info_t* p_job_info) {
     // TODO(bill_liu): implement
     (void)(job_id);
@@ -82,13 +79,25 @@ rdc_status_t RdcStandaloneHandler::rdc_job_get_stats(char  job_id[64],
     return RDC_ST_OK;
 }
 
-rdc_status_t RdcStandaloneHandler::rdc_job_stop_stats(char  job_id[64] ) {
+rdc_status_t RdcStandaloneHandler::rdc_job_stop_stats(char job_id[64]) {
     // TODO(bill_liu): implement
     (void)(job_id);
     return RDC_ST_OK;
 }
 
 
+rdc_status_t RdcStandaloneHandler::rdc_job_remove(char job_id[64]) {
+    // TODO(bill_liu): implement
+    (void)(job_id);
+    return RDC_ST_OK;
+}
+
+rdc_status_t RdcStandaloneHandler::rdc_job_remove_all() {
+    // TODO(bill_liu): implement
+    return RDC_ST_OK;
+}
+
+
 // Discovery RdcAPI
 rdc_status_t RdcStandaloneHandler::rdc_device_get_all(
         uint32_t gpu_index_list[RDC_MAX_NUM_DEVICES], uint32_t* count)  {