from msconfig.meta_config import MetaConfig
class FilenameIntroductionConfig(MetaConfig):
DATA = {
'timeline': [
('begin', "Timeline file description:\n"),
('msprof', "Timeline report.\n"
"\t1 CPU Layer: Data at the application layer, including the time "
"consumption information of upper-layer application operators. "
"The data needs to be collected only in msproftx or PyTorch scenarios.\n"
"\t2 CANN Layer: Data at the CANN layer, including the time consumption data of "
"components (such as AscendCL and Runtime) and nodes (operators).\n"
"\t3 Ascend Hardware Layer: Bottom-layer NPU data, including the time consumption data and "
"iteration trace data of each task stream under Ascend Hardware, "
"Communication and Overlap Analysis communication data, and other system data.\n"
"\t4 Overlap Analysis Layer: In cluster or multi-device scenarios, computation and "
"communication are sometimes parallel. You can check the pipeline overlapping time "
"(time when computation and communication are parallel) to determine "
"the computation and communication efficiencies.\n"
"\t\tCommunication Layer: Communication time.\n"
"\t\tCommunication(Not Overlaopped) Layer: Communication time that is not overlapped.\n"
"\t\tComputing: Computation time.\n"
"\t\tFree: Interval.\n"),
('step_trace', "You can determine the iteration that takes the longest time based on "
"the iteration length.\n"
"\tIteration ID: Iteration ID.\n"
"\tBP End: BP end time (ns).\n"
"\tFP_BP Time: FP/BP elapsed time (= BP End – FP Start). The unit is ns.\n"
"\tIteration Refresh: Iteration refresh hangover time (= Iteration End – BP End).\n"
"\tData_aug Bound: Data augmentation hangover time "
"(= Current FP Start – Previous Iteration End). The elapsed time of iteration 0 is "
"N/A because the previous Iteration End is absent.\n"
"\tReduce: Collective communication elapsed time (may involve groups of iterations). "
"If there is only one device, no Reduce data is output.\n"),
('msprof_tx', "msproftx timeline data. The collected host msproftx timeline data is combined "
"by thread for associated performance display.\n"
"\tevent_type: Event type.\n"),
],
'summary': [
('begin', "Summary file description:\n"),
('host_cpu_usage', "Host-side CPU utilization.\n"
"\tTotal Cpu Numbers: Total number of CPU cores in the system.\n"
"\tOccupied Cpu Numbers: Number of CPU cores occupied by processes.\n"
"\tRecommend Cpu Numbers: Number of CPU cores in use. In virtualization scenarios, "
"the value is the recommended number of CPU cores.\n"),
('host_mem_usage', "Host-side memory utilization.\n"
"\tTotal Memory: Total system memory (KB).\n"
"\tPeak Used Memory: Peak memory utilization (KB).\n"
"\tRecommend Memory: Recommended memory allocation in virtualization scenarios (KB).\n"),
('host_disk_usage', "Host-side disk I/O utilization.\n"
"\tPeak Disk Read: Recommended memory allocation in virtualization scenarios (KB).\n"
"\tRecommend Disk Read: Recommended memory allocation in "
"virtualization scenarios (KB).\n"
"\tPeak Disk Write: Recommended memory allocation in virtualization scenarios (KB).\n"
"\tRecommend Disk Write: Recommended disk write rate in "
"virtualization scenarios (KB/s)."),
('host_network_usage', "Host-side network I/O utilization.\n"
"\tNetcard Speed: NIC rated rate (KB/s).\n"
"\tPeak Used Speed: Maximum network rate (KB/s).\n"
"\tRecommend Speed: Recommended network rate in virtualization scenarios (KB/s).\n"),
('os_runtime_statistic', "Host-side syscall and pthreadcall.\n"
"\tCount: Number of calls to an API.\n"
"\tAvg, Max, Min\n: Average, maximum, and minimum duration of "
"the calls to an API (μs).\n"),
('cpu_usage', "System CPU usage on the host.\n"
"\tUser: Percentage of time taken to execute user-mode processes.\n"
"\tSys: Percentage of time taken to execute kernel-mode processes.\n"
"\tIoWait: Percentage of I/O wait time.\n"
"\tIrq: Percentage of hardware interrupt time.\n"
"\tSoft: Percentage of software interrupt time.\n"
"\tIdle: Percentage of idle time.\n"),
('process_cpu_usage', "Include host process cpu usage data.\n"),
('sys_mem', "Memory usage of processes on the host.\n"),
('process_mem', "System memory usage on the host.\n"),
('api_statistic', "Time spent by AscendCL API, is used to collect statistics on the time spent "
"by API execution at the CANN layer.\n"
"\tLevel: Level of an API, including AscendCL, Runtime, Node, Model, and"
" Communication.\n"),
('op_summary', "AI Core, AI CPU, AI Vector and COMMUNICATION communication operator data,"
"is used to collect statistics on operator details and time consumptions.\n"
"\tOp Name: Operator name.\n"
"\tOP Type: Operator type. \n"
"\tTask Type: Task type. \n"
"\tTask Start Time: Task start time (μs).\n"
"\tTask Duration: Task duration, including the scheduling time and the start time to "
"the latest end time of the first core. The unit is μs.\n"
"\tTask Wait Time: Interval between tasks (μs)."
"(= this task's start_time - last task's start_time - last task's duration_time)\n"
"\tBlock Dim: Number of running task blocks, which corresponds to the number of cores "
"during task running.\n"
"\tMix Block Dim: Number of running task blocks in Mix scenarios, which corresponds to "
"the number of cores during task running.\n"
"\tContext ID: Context ID.\n"
"\taiv_time: Average task execution duration on AIV."
"The value is calculated based on total_cycles "
"and mix block dim.\n"
"\taicore_time: Average task execution duration on AI Core."
"The value is calculated based on "
"total_cycles and block dim. The unit is μs. The data is inaccurate in "
"the manual frequency modulation, dynamic frequency modulation "
"(the power consumption exceeds the default value), and "
"Atlas 300V/Atlas 300I Pro scenarios. You are not advised referring to it.\n"
"\ttotal_cycles: Number of cycles taken to execute all task instructions.\n"
"\tRegister value: Value of the custom register whose data is to be collected.\n"),
('op_statistic', "AI Core and AI CPU operator calling times and time consumption.\n"
"The parameters of the msprof command line tool are used as examples. "
"The parameters of other collection modes are the same.Analyze the total calling time "
"and total number of calls of each type of operators, check whether there are "
"any operators with long total execution time, and analyze whether there is "
"any optimization space for these operators.\n"),
('step_trace', "Iteration trace data.\n"
"\tIteration End: End time of each iteration. The unit is μs.\n"
"\tIteration Time(us): Iteration time. (Iteration End of the current iteration - Iteration "
"End of the previous iteration). The Iteration End data of the previous iteration is "
"unavailable when the duration of the first iteration is calculated. "
"Therefore, Duration of the first iteration = Iteration End time of the current "
"iteration – FP start time of the current iteration. The unit is μs.\n"
"\tFP to BP Time(us): FP/BP elapsed time (= BP End – FP Start). The unit is μs.\n"
"\tIteration Refresh(us): Iteration refresh hangover time (= Iteration End – BP End). "
"The unit is μs.\n"
"\tData Aug Bound(us): Data augmentation hangover time "
"(= Current FP Start – Previous Iteration End). The elapsed time of iteration 0 "
"is N/A because the previous Iteration End is absent. The unit is μs.\n"
"\tModel ID: Graph ID in the model of a round of iteration.\n"
"\tReduce Start: Start time of collective communication.\n"
"\tReduce Duration(us): Total time spent by collective communication. "
"The collective communication duration is divided into two segments according to "
"the default segmentation policy. Reduce Start indicates the start time, and Reduce "
"Duration indicates the duration (μs) from the start to the end. Note that the Reduce "
"columns are not available in a single-device environment.\n"),
('communication_statistic', "Communication operator statistics, through which you can learn the time"
" consumption of this operator type and the time consumption ratio of each"
" Communication operator in collective communication to determine whether an"
" operator can be optimized.\n"),
('aicpu', "AI CPU summary.This file collects AI CPU data reported by data preprocessing. "
"Other files involving AI CPU data collect full AI CPU data.\n"),
('fusion_op', "Operator fusion summary.\n"),
('task_time', "Task Scheduler summary.\n"
"\tWaiting: Total wait time of a task (μs).\n"
"\tRunning: Total run time of a task (μs). An abnormally large value indicates that "
"the operator implementation needs to be improved.\n"
"\tPending: Total pending time of a task (μs).\n"),
('l2_cache', "L2 cache data.\n"
"\tHit Rate: L2 hit rate in requests received by DHA from AI Core\n"
"\tVictim Rate: Victim rate in requests received by DHA from AI Core.\n"),
('ai_core_utilization', "AI Core utilization.\n"
"\ticache_miss_rate: I-Cache miss rate. The smaller the value, "
"the higher the performance.\n"
"\tmemory_bound: AI Core memory bound, calculated as: "
"mte2_ratio/max(mac_ratio, vec_ratio). If the value is less than 1, "
"no memory bound exists. If the value is greater than 1, a memory bound exists. "
"The greater the value, the severer the bound.\n"),
('ai_vector_core_utilization', "Percentage of instructions on each AI Vector Core.\n"
"\ticache_miss_rate: I-Cache miss rate. The smaller the value, "
"the higher the performance.\n"
"\tmemory_bound: AI Core memory bound, calculated as: "
"mte2_ratio/max(mac_ratio, vec_ratio). If the value is less than 1, "
"no memory bound exists. If the value is greater than 1,"
"a memory bound exists. "
"The greater the value, the severer the bound.\n"),
('memory_record', "Records the memory applied for by the GE component at the CANN level"
"and the memory usage time.\n"),
('operator_memory', "Records the memory required and occupied time when"
"CANN level operators are executed on the NPU. The memory is applied by the GE.\n"),
('ddr', "DDR memory read/write speed.\n"),
('hbm', "HBM memory read/write speed.\n"),
('npu_mem', "NPU memory usage data.\n"
"\tevent: Event type.device app\n"
"\tddr(KB): DDR memory usage. This field is not supported by the "
"current product and is displayed as 0.\n"
"\thbm(KB): HBM memory usage. This field is not supported by the "
"current product and is displayed as 0.\n"
"\tmemory(KB): Sum of DDR and HBM memory usages.\n"),
('hccs', "HCCS bandwidth.\n"
"\tMode: Tx and Rx: receive bandwidth and transmit bandwidth (MB/s).\n"),
('roce', "RoCE bandwidth.\n"
"\tRx Bandwidth efficiency: Receive bandwidth.\n"
"\trxPacket/s: Packet receive rate.\n"
"\trxError rate: Received packet error rate.\n"
"\trxDropped rate: Received packet loss rate.\n"
"\tTx Bandwidth efficiency: Transmit bandwidth.\n"
"\ttxPacket/s: Packet transmit rate.\n"
"\ttxError rate: Transmitted packet error rate.\n"
"\ttxDropped rate: Transmitted packet loss rate.\n"
"\tfuncId: Node ID.\n"),
('pcie', "PCIe bandwidth.\n"),
('nic', "NIC summary.\n"
"\tRx Bandwidth efficiency: Receive bandwidth.\n"
"\trxPacket/s: Packet receive rate.\n"
"\trxError rate: Received packet error rate.\n"
"\trxDropped rate: Received packet loss rate.\n"
"\tTx Bandwidth efficiency: Transmit bandwidth.\n"
"\ttxPacket/s: Packet transmit rate.\n"
"\ttxError rate: Transmitted packet error rate.\n"
"\ttxDropped rate: Transmitted packet loss rate.\n"
"\tfuncId: Node ID.\n"),
('dvpp', "DVPP data.\n"
"\tDvpp Id: DVPP ID.The ID range is related to the number of devices on the AI Server. "
"Each device is mapped to five IDs, indexed starting at 0.\n"),
('llc_read_write', "L3 cache read/write speed.\n"
"\tHit Rate: L3 cache hit rate.\n"
"\tThroughput: L3 cache throughput (MB/s).\n"),
('llc_aicpu', "L3 cache used by AI CPU.\n"
"\tUsed Capacity of LLC: Used L3 cache (MB).\n"),
('llc_ctrlcpu', "L3 cache used by Ctrl CPU.\n"
"\tUsed Capacity of LLC: Used L3 cache (MB).\n"),
('llc_bandwidth', "L3 cache bandwidth."),
('ai_cpu_top_function', "AI CPU top functions.\n"
"\tFunction: Top functions of AI CPU.\n"
"\tModule: Module where the function is located.\n"
"\tCycles: Cycles taken to execute a function in the sampling period.\n"
"\tCycles(%): Percentage of cycles taken to execute a function "
"in the sampling period.\n"),
('ai_cpu_pmu_events', "AI CPU PMU events.\n"),
('ctrl_cpu_pmu_events', "Ctrl CPU top functions.\n"
"\tFunction: Top functions of Ctrl CPU.\n"
"\tModule: Module where the function is located.\n"
"\tCycles: Cycles taken to execute a function in the sampling period.\n"
"\tCycles(%): Percentage of cycles taken to execute a function "
"in the sampling period.\n"),
('ctrl_cpu_pmu_events', "Ctrl CPU PMU events.\n"),
('ts_cpu_top_function', "TS CPU top functions.\n"
"\tFunction: Top functions of TS CPU.\n"
"\tCycles: Cycles taken to execute a function in the sampling period.\n"
"\tCycles(%): Percentage of cycles taken to execute a function "
"in the sampling period.\n"),
('ts_cpu_pmu_events', "TS CPU PMU events.\n"),
('msprof_tx', "msproftx summary data. The collected host msproftx summary data is combined by "
"thread for associated performance display.\n"
"\tmessage: Character string description carried in the msproftx profiling process.\n"),
]
}
def __init__(self):
super().__init__()
