Update slides

application-performance/docs/02-gpu-performance.md

@@ -1,27 +1,30 @@
 ---
-title:  Single node performance optimization
+title:  GPU performance optimization
 event:  CSC Summer School in High-Performance Computing 2024
 lang:   en
 ---
 
 # GPU performance optimization {.section}
 
 # Introduction
-- GPUs (Graphics Processing Units) are widely used in High-Performance Computing (HPC) applications.
-- GPUs are powerful and complex processors designed for parallel computing.
-- GPUs require explicit expression of parallelism by the programmer.
+
+- GPUs (Graphics Processing Units) are widely used in High-Performance Computing (HPC) applications
+- GPUs are powerful and complex processors designed for parallel computing
+- GPUs require explicit expression of parallelism by the programmer
 
 # General Principles for High GPU Performance 
 
-<div class=column>
+:::::: {.columns}
+::: {.column width="50%"}
 - Keep all the compute resources busy (idle resources are a waste)
 - Minimize the synchronization at all levels
 - Minimize the data transfers between host and device 
 - Keep the data in faster memory and use an appropriate access pattern
-</div>
-<div class=column>
-![](img/lumi_node.png){.center width=40%}
-</div>
+:::
+::: {.column width="50%"}
+![](img/lumi_node.png){.center width=50%}
+:::
+::::::
 
 # GPU performance analysis {.section}
 
@@ -30,6 +33,7 @@ lang:   en
 ![](img/perf-analysis-single-gpu.svg){.center width=60%}
 
 # Measuring performance
+
 - Don’t speculate about performance – measure it!
 - Performance analysis tools help to
     - Find hot-spots
@@ -43,12 +47,10 @@ lang:   en
 
 # Hardware performance counters
 
-- Hardware performance counters are special registers on CPU \& GPU that count
-  hardware events
-- They enable more accurate statistics and low overhead
+- Special registers on CPU \& GPU that count hardware events
+- Enable more accurate statistics and low overhead
     - In some cases they can be used for tracing without any extra
       instrumentation
-
 - Number of counters is much smaller than the number of events that can be
   recorded
 - Different devices have different counters
@@ -71,71 +73,65 @@ lang:   en
 - Start with an overview!
     - Call tree information, what routines are most expensive?
 
-# Sampling vs. Tracing
+# Sampling
 
-<div class=column>
-Sampling
+:::::: {.columns}
+::: {.column width="50%"}
 
 - Application is stopped at predetermined intervals 
     - Information is collected about the state of application
 - Lightweight, but may give skewed results
     - Statistical information
 
-</div>
-<div class=column>
-Tracing
+:::
+::: {.column width="50%"}
 
-- Records events, e.g., every function call
-- Requires usually modification to the executable *i.e.* instrumentation
-- More accurate, but may affect program behavior
-- Generates often lots of data
-
-</div>
+![](img/sampling.png){.left width=100%}
 
-# Sampling vs. Tracing
+:::
+::::::
 
-<div class=column>
-Sampling
+# Tracing
 
-![](img/sampling.png){.left width=80%}
+:::::: {.columns}
+::: {.column width="50%"}
 
-</div>
-<div class=column>
-Tracing
+- Records events, e.g., every function call
+- Requires usually modification to the executable: code instrumentation
+- More accurate than sampling, but may affect program behavior
+- Generates often lots of data
 
-![](img/tracing.png){.left width=80%}
+:::
+::: {.column width="50%"}
 
-</div>
+![](img/tracing.png){.left width=100%}
 
-# Tau  Analysis Utilities
+:::
+::::::
 
-<small> 
+# Tau Analysis Utilities
 
-- TAU is a powerful performance evaluation toolkit
+- <https://www.cs.uoregon.edu/research/tau/home.php>
+- A performance evaluation toolkit
 - Runs on all HPC platforms, relatively easy to install
 - Targets all parallel programming/execution paradigms (GPU, MPI, OpenMP, pthreads, ...)
-- Programming languages: Fortran, C, C++, UPC,  Java, Python, ...
+- Programming languages: Fortran, C, C++, UPC, Java, Python, ...
+
+# Tau Analysis Utilities cont.
+
 - TAU has instrumentation, measurement and analysis tools 
 - User-friendly graphical interface
 - Profiling: Measures total time spent in each routine
 - Tracing: Shows events and their timings across processes on a timeline
 - I/O performance evaluation
 - Memory debugging
 
-</small>
-
 # Omniperf Tools
 
+- <https://amdresearch.github.io/omniperf/getting_started.html>
  - system performance profiling tool for machine
-learning/HPC workloads running on AMD MI GPUs. 
- - presently targets usage on MI100 and MI200 accelerators.
+learning/HPC workloads running on AMD MI GPUs 
+ - presently targets usage on MI100 and MI200 accelerators
  - profiling, roofline model, tracing
  - built on top of `roctracer` and `rocprof`  
- - supports both a web-based GUI and a command-line analyzer for user convenience.
-
-
-# Web Resources
- - TAU homepage
-    - <https://www.cs.uoregon.edu/research/tau/home.php>
- - Omniperf
-    - <https://amdresearch.github.io/omniperf/getting_started.html>
+ - supports both a web-based GUI and a command-line analyzer for user convenience