The purpose of this lab is the familiarization with the field of application profiling & debugging, through the means of dedicated tools for spotting performance bottlenecks.

We will focus on the following open source tools:

• valgrind (callgrind, cachegrind) with callgrind_annotate and cg_annotate for results visualization (on haswell partition)
• perf (on your own systems only, as hardware counters are not available inside Apptainer containers)

Valgrind is a tool used for memory debugging, memory leak detection and profiling. It is also a generic framework for creating dynamic analysis tools, such as memory checkers [1].

Valgrind is in essence a virtual machine using just-in-time compilation techniques, including dynamic recompilation. It is important to keep in mind that nothing from the original program ever gets run directly on the host processor. Instead, it will translate the input program into a simpler form called Intermediate Representation (IR), which is processor neutral. After this transformation, a tool [2] is called to do whatever transformation of the IR it needs and the resulting IR is then translated back into machine code and ran on the host processor.

The tools available in Valgrind are:

• memcheck. Used to detect memory-management problems and it is aimed at C and C++ programs. All memory reads and writes are checked, and all calls to malloc/new/free/delete are intercepted. Therefore it can detect memory leaks, access to invalid memory, weird initialization values, overflows, etc. Memcheck runs programs about 10-30x slower than normal;
• cachegrind. Used to profile CPU cache. It performs detailed simulation of the I1, D1 and L2 caches in order to pinpoint the sources of cache misses. It identifies the number of cache misses, memory references and instructions executed for each line of source code. Cache grind runs programs about 20-100x slower than normal;
• callgrind. It is an extension to cachegrind and provides all the information that the latter offers, plus extra information regarding call graphs. In order to view the results, a visualization tool called KCachegrind [3] can be used;
• massif. It is a heap profiler and it performs detailed heap profiling by taking regular snapshots of a program's heap and produces a graph showing heap usage over time, including information about which parts of the program are responsible for the most memory allocations. Massif runs programs about 20x slower than normal;
• helgrind and drd. These tools are thread debuggers which find data races in multithreaded programs. They look for memory locations which are accessed by more than one (POSIX) pthread, but for which no consistently used (pthread_mutex_) lock can be found;
• other 3rd party tools can be found here [4].

In this section, we will focus on analyzing a software application. We will analyze both a serial and a parallel implementation. The application is called “tachyon” and you can find the source code attached to this lab.

On your own system, before compilation, you must install the X11 dev tools and create a set of symlinks. For Ubuntu 64 bit, we must do the following:

• install dependencies

   sudo apt-get install libx11-dev 

• create the symlinks:

  •  sudo mkdir /usr/lib64 

  •  sudo ln -s /usr/lib/x86_64-linux-gnu/libX11.so /usr/lib64/libX11.so 

  •  sudo ln -s /usr/lib/x86_64-linux-gnu/libXext.so /usr/lib64/libXext.so 

To compile it, you must extract the archive Tachyon to local disk and run make. You can test the compilation by running in the same directory:

./tachyon_find_hotspots dat/balls.dat 

You should see a window like the one below:

To run and observe the functionalities of this tool, follow the following sequence of steps and instructions:

wget -O tachyon_vtune_amp_xe.tgz http://ocw.cs.pub.ro/courses/_media/asc/lab6/tachyon_vtune_amp_xe.tgz 
gunzip tachyon_vtune_amp_xe.tgz
tar -xvf tachyon_vtune_amp_xe.tar
cd tachyon
make

1. Make sure you have Valgrind and KCachegrind installed on the system (or login on the hp-sl.q queue) and the application in the initial state, without any modifications on your system

sudo apt-get update
sudo apt-get install valgrind kcachegrind

2. We will use the tool callgrind to get information from the running application. Run the following command line:

valgrind --tool=callgrind --collect-jumps=yes --dump-instr=yes --collect-systime=yes -- ./tachyon_find_hotspots dat/balls.dat

3. Open the profile in KCachegrind and click on the Calee Map tab. Also, make sure that the buttons % Relative, Cycle detection and Relative to parent are selected. You should see something like this: From this image, we can see that valgrind measured that about 78% of the total time was spent in the initialize_2D_buffer function. Double click the square containing the function name, then select the “Source code” tab and you will see the problematic code.

Perf is a performance analysis tool, available in the Linux kernel since version 2.6.31 [5]. The userspace control application is accessed from the command line and provides a number of subcommands. Unlike Valgrind, perf is capable of statistical profiling of both the entire system (kernel and userspace) and per process PID basis. It supports hardware performance counters, tracepoints, software performance counters (e.g. hrtimer), and dynamic probes (for example, kprobes or uprobes).

Perf is used with several subcommands:

• stat: measure total event count for a single program or for the whole system for a specified time period;
• top: top-like dynamic view of hottest functions;
• record: measure and save sampling data for single program;
• report: analyze file generated by perf record;
• annotate: annotate sources or assembly;
• sched: tracing/measuring of scheduler actions and latencies;
• list: list available events.

1. Make sure you have perf installed on the system and the application in the initial state, without any modifications. You can only run perf as root. You can only do this on your system. 2. Run the following command line:

perf record -a -g -- ./tachyon_find_hotspots

For other perf parameters, you can read this link 3. Run the following command line to view the collected results:

perf report

You should see a screen like the following: From this image you can see that perf will display the symbol for the function that takes the most amount of CPU time in red. In our case it’s the _Z20initialize_2D_bufferPjS_, which translates in the C source code into the same function as with VTune and Valgrind.

 c++filt _Z20initialize_2D_bufferPjS_

Pentru acest laborator se pot utiliza sistemele din cluster prin intermediul fep.grid.pub.ro:

1. Vă conectați pe fep, conform Laboratorului 1 (Vezi Ce este FEP-ul?).
2. Downloadați arhiva laboratorului: wget https://ocw.cs.pub.ro/courses/_media/asc/laboratoare/lab4_skl.tar.gz -O lab4_skl.tar.gz.
3. Dezarhivați arhiva downloadată mai sus: tar -xzvf lab4_skl.tar.gz.
4. Intrați în mod interactiv pe Haswell: srun --pty bash.
5. Accesați imaginea de docker în cadrul căreia veți face laboratorul:apptainer run docker://gitlab.cs.pub.ro:5050/asc/asc-public/c-labs:1.3.1 /bin/bash

Task 0 - Folositi Callgrind pentru task0.c, urmărind TODO-urile din cod.

# Versiunea serială
Apptainer> make task0
Apptainer> valgrind --tool=callgrind -v --dump-every-bb=10000000 ./task0
Apptainer> callgrind_annotate callgrind.out.<pid>
 
# Versiunea paralelizată
Apptainer> make clean
Apptainer> make openmp_task0
Apptainer> valgrind --tool=callgrind -v --dump-every-bb=10000000 ./task0
Apptainer> callgrind_annotate callgrind.out.<pid>

Notați și explicați următoarele observații:

• Care este funcția cu cel mai mare număr de instrucțiuni (coloana Ir)?
• Cum diferă profiling-ul serial de cel paralel? Ce funcții noi apar în versiunea paralelizată?
• De ce ordinea primelor 100 de prime diferă între versiunea serială și cea paralelă?

Task 1 - Analizati aplicatia Tachyon (ray tracer).

Apptainer> ./task1.sh
 
# Versiunea serială
Apptainer> valgrind --tool=callgrind --collect-jumps=yes --dump-instr=yes --collect-systime=yes -- ./tachyon_find_hotspots dat/balls.dat
Apptainer> callgrind_annotate callgrind.out.<pid>
 
# Versiunea paralelizată
Apptainer> valgrind --tool=callgrind --collect-jumps=yes --dump-instr=yes --collect-systime=yes -- ./tachyon_analyze_locks dat/balls.dat
Apptainer> callgrind_annotate callgrind.out.<pid>
 
# Perf (doar software counters în container)
Apptainer> perf stat ./tachyon_find_hotspots dat/balls.dat
Apptainer> perf stat ./tachyon_analyze_locks dat/balls.dat

Notați și explicați următoarele observații:

• Care sunt funcțiile cu cel mai mare număr de instrucțiuni în versiunea serială?
• Ce funcție consumă cel mai mult din versiunea serială și de ce? (hint: initialize_2D_buffer)
• Cum diferă profiling-ul serial de cel paralel? Ce funcții noi apar?
• Unde este plasat mutex-ul în analyze_locks? Este necesar acolo?

Task 2 - Analizați înmulțirea de matrice cu diferite ordonări ale buclelor folosind Cachegrind.

• Compilati si rulati task2.c
• Notati observatiile voastre legate de numarul de I refs, D refs, D1 misses, branches si mispredicts.

Apptainer> make task2
Apptainer> valgrind --tool=cachegrind --cache-sim=yes ./task2 1
Apptainer> valgrind --tool=cachegrind --cache-sim=yes ./task2 2
Apptainer> valgrind --tool=cachegrind --cache-sim=yes ./task2 3

Completați tabelul cu rezultatele obținute:

Notați și explicați următoarele observații:

• Comparați valorile I refs și D refs între cele 3 ordonări. Ce observați?
• Comparați valorile D1 miss rate între cele 3 ordonări. Care mod este mai eficient din punct de vedere al cache-ului și de ce?
• Comparați valorile LLd miss rate între cele 3 ordonări. Ce observați?
• Cum influențează ordinea buclelor accesul la memorie în C?

• http://valgrind.org/
• http://valgrind.org/info/tools.html
• http://kcachegrind.sourceforge.net/html/Usage.html
• http://valgrind.org/downloads/variants.html
• https://perf.wiki.kernel.org/index.php/Main_Page
• https://software.intel.com/en-us/intel-parallel-studio-xe
• http://www.brendangregg.com/perf.html
• https://www.oracle.com/tools/developerstudio/downloads/developer-studio-jsp.html
• https://guides.upb.ro/docs/grid/apptainer-usage

• Responsabilii acestui laborator: Emil Slusanschi, Alexandru Patrascu si Octavian Moraru.
• PDF laborator
• Aplicatie laborator (tachyon)
• Schelet laborator