add full mpi metric tooltips

Author: ptooley

pypop/metrics/mpi.py

@@ -10,22 +10,82 @@
 
 from .metricset import MetricSet, Metric
 
+k_GE_desc = (
+    "The overall quality of the parallelisation.  This is the product of the "
+    "Parallel Efficiency and Computational Scaling"
+)
+k_PE_desc = (
+    "The overall efficiency with which the computation is parallelised between "
+    "different processes and threads. This is further divided into the Process Level"
+    "Efficiency and the Thread Level Efficiency"
+)
+k_PLE_desc = (
+    "The efficiency of the application as viewed at the process level, including the "
+    "MPI communication and process-level load balance."
+)
+k_MPILB_desc = (
+    "The efficiency with which the total amount of computational work is shared "
+    "between the different MPI processes. Low values indicate that there is significant "
+    "imbalance between the most and least loaded processes."
+)
+k_MPICE_desc = (
+    "The efficiency with which the application carries out MPI communication.  An ideal "
+    "application will spend no time communicating and 100% of time in computation. Low "
+    "values indicate that too much communication is being performed for the amount of "
+    "computation."
+)
+k_MPITE_desc = (
+    "The efficiency of the actual transfer of data via MPI. This reflects the size of "
+    "the data being communicated and the speed of the underlying communication network. "
+    "Low values indicate that the network bandwidth is insufficient for the required "
+    "communication rate, or that too much data is being communicated."
+)
+k_MPISE_desc = (
+    "The efficiency with which the MPI communications are synchronised and carried out. "
+    "Low values indicate that there is significant irregularity in the timings of "
+    "different processes arrivals at MPI calls, reducing efficiency due to waiting "
+    "for MPI calls to complete."
+)
+k_COMPSC_desc = (
+    "The way in which the total computational cost varies with the applied parallelism. "
+    "This is a combination of the increased cost due to additional calculations "
+    "performed, and increased costs due to reduced instructions per cycle."
+)
+k_INSSC_desc = (
+    "Inefficiencies introduced due to an increase in the total computational work done, "
+    "measured by the total CPU instructions. Ideally, there would be no additional "
+    "computation required when parallelising, but there is normally some additional "
+    "cost to manage the distribution of work. The Instruction Scaling metric "
+    "represents this by calculating the relative difference in total instructions "
+    "between runs."
+)
+k_IPCSC_desc = (
+    "Inefficiencies due to changes in the instructions per cycle executed by the CPUs. "
+    "The IPC rate can be reduced due to CPU data starvation, inefficient cache usage or "
+    "high rates of branch misprediction."
+)
+k_FREQSC_desc = (
+    "Inefficiencies due to changes in the rate at which the CPU executes instructions. "
+    "This is typically due to thermal management in the CPU reducing the overall clock "
+    "speed."
+)
+
 
 class MPI_Metrics(MetricSet):
     """Pure MPI Metrics  (additive version).
     """
 
     _metric_list = [
-        Metric("Global Efficiency", 0),
-        Metric("Parallel Efficiency", 1),
-        Metric("MPI Load Balance", 2, "Load balance"),
-        Metric("MPI Communication Efficiency", 2),
-        Metric("MPI Transfer Efficiency", 3),
-        Metric("MPI Serialisation Efficiency", 3),
-        Metric("Computation Scaling", 1),
-        Metric("Instruction Scaling", 2),
-        Metric("IPC Scaling", 2),
-        Metric("Frequency Scaling", 2),
+        Metric("Global Efficiency", 0, desc=k_GE_desc),
+        Metric("Parallel Efficiency", 1, desc=k_PE_desc),
+        Metric("MPI Load Balance", 2, "Load balance", desc=k_MPILB_desc),
+        Metric("MPI Communication Efficiency", 2, desc=k_MPICE_desc),
+        Metric("MPI Transfer Efficiency", 3, desc=k_MPITE_desc),
+        Metric("MPI Serialisation Efficiency", 3, desc=k_MPISE_desc),
+        Metric("Computation Scaling", 1, desc=k_COMPSC_desc),
+        Metric("Instruction Scaling", 2, desc=k_INSSC_desc),
+        Metric("IPC Scaling", 2, desc=k_IPCSC_desc),
+        Metric("Frequency Scaling", 2, desc=k_FREQSC_desc),
     ]
 
     _programming_model = "MPI"