Differences

This shows you the differences between two versions of the page.

--- skill-tree:pe:2:2:b [2020/06/18 20:15] – external edit 127.0.0.1
+++ skill-tree:pe:2:2:b [2025/04/16 18:30] (current) – external edit 127.0.0.1
@@ Line 1: / Line 1: @@
-# PE2.2-B Profiling
+# PE2.2 I/O Performance
-# Background
-  *  Profiling is explained for the CPU level, where it can be supported by hardware performance counters and by sampling techniques.
+There are several aspects involved in delivering high I/O performance to parallel applications, from hardware characteristics to methods that manipulate workloads to improve achievable performance.
-  *  Sampling is used to see, by examining the program counter, what routines and source code lines of a program are responsible for which portions of the total runtime.
-  *  Automatically adding trace code to a parallel program by so-called instrumentation to record its execution in a strict chronology is explained and the difference to profiling is emphasized.
+Running the same application with different I/O configurations gives the possibility to tune the I/O system according to the application access pattern.
-  *  Similar techniques are explained for profiling the network level (e.g. based on InfiniBand counters and I/O server states).
-# Aim
+One way to predict application performance in HPC systems with different I/O configurations is by using modelling and simulation techniques.
-# Outcomes
+Modeling the system allows assessing obtained performance and therewith estimate the performance potentially gained by optimizations.
-  *  detect performance issues and bottlenecks caused, for example, by inefficient programming, memory accesses, I/O operations, cache-misses, page-faults, and parallelization overheads
-  *  use environment variables like $I_MPI_STATS to control the built-in performance analysis functionality in MPI
+File systems are implemented in the operating system which deploys strategies to improve performance such as scheduling, caching and aggregation.
+Therefore, the observable I/O performance depends on more than the capabilities of the raw block device.
+## Learning Outcomes
+* Select performance models to assess and optimize the application I/O performance.
+* Identify tools capable of predicting the behavior of applications in HPC.
+* Apply methods to manipulate workloads to improve achievable performance.
+* Understand metrics that describe relevant I/O performance characteristics.
+* Select tools to analyze I/O performance of applications.
+* Apply the performance analysis tools to applications and workflows.
+* Interpret the results of I/O analysis tools and distill optimization strategies for the application or system configurations.
-# Subskills
-  * [[skill-tree:pe:2:2:1:b]]