diff --git a/README.md b/README.md index fa23e29..12a273e 100644 --- a/README.md +++ b/README.md @@ -2,20 +2,51 @@ **Important**: these are instructor notes, remove this file before showing the materials to the students. The notes can be added after the lecture, of course. ## Introduction - - [Puzzle](puzzle.ipynb) (how swapping two nested for-loops makes out for a >27× slowdown + - [Puzzle](/puzzle.ipynb) ➔ [read-only rendered notebook](https://nbviewer.org/urls/git.aspp.school/ASPP/2024-heraklion-comp-arch/raw/branch/main/puzzle.ipynb) + - Question: how come that swapping dimensions in a for-loop makes out for a huge slowdown? - Let students play around with the notebook and try to find the "bug" - - A more thorough benchmark using the same code is [here](benchmark_python/) + - A more thorough [benchmark](benchmark_python/) + ## A digression in CPU architecture and the memory hierarchy - - Go to [A Primer in CPU architecture](architecture) - - The need for a hierarchical access to data for the CPU should be clear now ➔ the "starving" CPU problem - - Have a look at the historical evolution of [speeds](speed/) of different components in a computer: - - the CPU clock rate - - the memory (RAM) bandwidth, latency clock rate - - the storage media access rates + - Go to [A Primer in CPU architecture](architecture/) + - Measure size and timings for the memory hierarchy on my machine with a low level [C benchmark](benchmark_low_level/) + +## Analog programming +Two exercises to activate the body and the mind + +Common goal of both exercises is to sort a deck of tarot cards by value +### First experiment: human sorting +Setup: +- 1 volunteer to keep the time spent sorting +- each person picks up a tarot card from the randomly shuffled deck on the table +- moving around and speaking is allowed until the tarot cards are displayed sorted on the table + +### Second experiment: machine sorting +Setup: +- 2 volunteers to keep the time: + - one volunteer keeps the time spent *programming* + - one volunteer keeps the time spent *executing* the program +- 2 volunteers to be the *programmers*: + - can use the whiteboard + - can and should speak and think loudly and ask for help +- 2 volunteers to be two CPUs: + - only understand the instructions: + - **fetch** a value from a memory address into register `N` ➔ returns `0` if succeded else `1` + - **push** the value from register `N` to a memory address ➔ returns `0` if succeded else `1` + - **compare** var0 and var1 ➔ returns `0` if `var0 ≥ var1` else `1` +- 4 volunteers to be CPU registers: + - each register has a tag: `R1`, `R2`, `R3`, `R4` + - a value fetched from memory is kept in short-term memory by the registers + - the result value of an operation is stored in one register +- everyone else sits on their seats and represent RAM: + - they own a *value*, i.e. they hold on a tarot card + - they have an address based on their seating order: 0th seat, 1st seat, 2nd seat, 3rd seat, 4th seat, etc… + - when *fetched*, walk to the corresponding register and hand in their *value* (card) + - when *pushed*, walk to the corresponding register and fetch their new *value* (card) +- each RAM address comes and picks up a random tarot card as initialization step - - Measure size and timings for the memory hierarchy on my machine with a low level [C benchmark](benchmark_low_level) ## Back to the Python benchmark (second try) @@ -25,19 +56,28 @@ ## Anatomy of a numpy array - - [memory layout of numpy arrays](numpy) + - [memory layout of numpy arrays](numpy/) ## Back to the Python benchmark (third try) - can we explain what is happening now? Yes, more or less ;-) - - quick fix for the [puzzle](puzzle.ipynb): try and add `order='F'` in the "bad" snippet and see that is "fixes" the bug ➔ why? + - quick fix for the [puzzle](/puzzle.ipynb): try and add `order='F'` in the "bad" snippet and see that is "fixes" the bug ➔ why? Notes on the [Python benchmark](benchmark_python/): - while running it attached to the P-core (`cpu0`), the P-core was running under a constant load of 100% (almost completely user-time) and at a fixed frequency of 3.8 GHz, where the theoretical max would be 5.2 GHz - - while running it attached to the E-core (`cpu10`), the E-core was running under a constant load of 100% (almost completely user-time) and at a fixed requency of 2.5 GHz, where the theoretical max would be 3.9 GHz - - ... ➔ the CPU does not "starve" because it scales its speed down to match the memory throughput? Or I am misinterpreting this? This problem which at first sight should be perfectly memory-bound, becomes CPU-bound, or actually, exactly balanced? ;-) - -## Excerpts of parallel Python - - [The dangers and joys of automatic parallelization](parallel) (like in numpy linear algebra routines) and the use of clusters/schedulers (but also on your laptop) + + ➔ the CPU does not "starve" because it scales its speed down to match the memory throughput? Or I am misinterpreting this? This problem which at first sight should be perfectly memory-bound, becomes CPU-bound, or actually, exactly balanced? From the [Intel documentation](https://lenovopress.lenovo.com/lp1836-tuning-uefi-settings-4th-gen-intel-xeon-scalable-processor): + > **Energy Efficient Turbo** + > + > When `Energy Efficient Turbo` is enabled, the CPU’s optimal turbo + > frequency will be tuned dynamically based on CPU utilization. The actual + > turbo frequency the CPU is set to is proportionally adjusted based on the + > duration of the turbo request. Memory usage of the OS is also monitored. + > If the OS is using memory heavily and the CPU core performance is limited + > by the available memory resources, the turbo frequency will be reduced + > until more memory load dissipates, and more memory resources become + > available. The power/performance bias setting also influences energy + > efficient turbo. `Energy Efficient Turbo` is best used when attempting to + > maximize power consumption over performance. ## Concluding remarks - how is all of this relevant for the users of a computing cluster?