Hardware: Jûrgen and others make a very good point about the memory interface channel speed to the CPUs as a factor that SW cannot alter.
Software speedups: Back in the “old” days when I worked at I.P. Sharp we did make some hand crafted speedups in the interpreter to handle special cases. I ws tasked with looking into what could be done with inner and outer product dealing with large arrays. We did get a significant speed up for those special cases. Other areas were transpose and grade up/down again for large many dimensional arrays. But that was IBM Mainframes running DOS and later MFT/MVT.
In today’s computer world, specially for desktop use, we can get multi-CPU multiple core boxes to use. For example have a look at some of the Raspberry Pi clusters that have been built. However the memory to CPU channel is still the primary bottleneck, coupled with the lack of massive, affordable on-board “RAM”. Lucky it’s not “CORES” as in my days.
Maybe of some value is finding the optimal algorithms, in the interpreter, for searching, sorting, transposing, multiplying massive multiple many dimensional data bases. And also a way to up the priority in the OS’ dispatcher for the interpreter and it APL related tasks. But that is not a simple matter given virtual memory..
Just the ravings of an old fool….
Respect…
Peter
CPU cache is the reason for a JIT compilation/conversion of APL expressions, in order to fuse local operations together. If you are in to CPU cache aspects of HPC, the blis papers would be interesting to you. See the citations at https://github.com/flame/blisOn Oct 17, 2019, at 10:03 AM, Rowan Cannaday <address@hidden> wrote:
Peter:
I am new to APL (~9 months). Most of my day-to-day work is sql/shell, however I use APL for a couple things: 1.) as an ad-hoc calculator & 2.) a symbolic notation that greatly simplifies complex mathematical calculations in how I think about, remember, and approach them.
As for things I've built in APL, its been mostly toy projects. I built a simple artificial neural net, however ran into difficulties when attempting to generalize backpropagation for arbitrary array sizes. I would like to complete this at some point.
Parallelizing APL was more of a curiosity, than an immediate need. I do not plan to invest much energy into pursuing it at this time (especially in light of Jürgen's explanation). To be honest I had watched a talk about cpu caching in C++ and was interested in how that related to APL's handling of arrays.
On Wed, Oct 16, 2019 at 9:53 PM Peter Teeson <address@hidden> wrote: Hi Rowan:
What classes of problems are you trying to solve that would benefit from parallel processing?
Respect
Peter Teeson
On Oct 16, 2019, at 1:27 PM, Dr. Jürgen Sauermann <mail@xn--jrgen-sauermann-zvb.de> wrote:
Hi Rowan,
actually there is no syntax tree in GNU APL. The way in which APL binds names (*late and ambiguously) makes it fairly useless to parse it beforehand. What happens in GNU APL is prefix matching at runtime. The prefix-table is in src/Prefix.def (an automatically generated hast table that does lookups essentially in time O(1) per prefix.)
This lookup table replaces the AST that you would have in a compiled language,
Best regards, Jürgen Sauermann
On 10/16/19 6:55 PM, Rowan Cannaday wrote:
Thanks again,
AST = abstract syntax tree. The tree-like structure that is produced by the parser.
Avoiding compilation is a reasonable restriction.
Thanks for the context.
- Rowan
``` #gdb apl GNU gdb (Debian 8.3.1-1) 8.3.1 Copyright (C) 2019 Free Software Foundation, Inc. License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html> This is free software: you are free to change and redistribute it. There is NO WARRANTY, to the extent permitted by law. Type "show copying" and "show warranty" for details. This GDB was configured as "x86_64-linux-gnu". Type "show configuration" for configuration details. For bug reporting instructions, please see: <http://www.gnu.org/software/gdb/bugs/>. Find the GDB manual and other documentation resources online at: <http://www.gnu.org/software/gdb/documentation/>.
For help, type "help". Type "apropos word" to search for commands related to "word"... Reading symbols from apl... (gdb) run Starting program: /usr/local/bin/apl [Thread debugging using libthread_db enabled] Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1". [Detaching after vfork from child process 23377] [New Thread 0x7ffff68d7700 (LWP 23381)] [New Thread 0x7ffff60d6700 (LWP 23382)] [New Thread 0x7ffff58d5700 (LWP 23383)]
______ _ __ __ __ ___ ____ __ / ____// | / // / / / / | / __ \ / / / / __ / |/ // / / / / /| | / /_/ // / / /_/ // /| // /_/ / / ___ | / ____// /___ \____//_/ |_/ \____/ /_/ |_|/_/ /_____/
Welcome to GNU APL version 1.8 / 1191M
Copyright (C) 2008-2019 Dr. Jürgen Sauermann Banner by FIGlet: www.figlet.org
This program comes with ABSOLUTELY NO WARRANTY; for details run: /usr/local/bin/apl --gpl.
This program is free software, and you are welcome to redistribute it according to the GNU Public License (GPL) version 3 or later.
)OFF
Goodbye. Session duration: 5.06809 seconds Couldn't read debug register: No such process. (gdb) Cannot find user-level thread for LWP 23382: generic error (gdb) [Thread 0x7ffff58d5700 (LWP 23383) exited] [Thread 0x7ffff60d6700 (LWP 23382) exited] [Thread 0x7ffff68d7700 (LWP 23381) exited] [Inferior 1 (process 23368) exited normally]
(gdb) bt No stack.
```
On Wed, Oct 16, 2019 at 4:18 PM Dr. Jürgen Sauermann <mail@jürgen-sauermann.de> wrote: Hi Rowan,
a stack-trace for the segfault would be good (command gdb apl then: 'run' and finally 'bt' after the segfault,
No idea what AST is. You could try TAB-expansion to get options in various situations and try e.g.
]help ⌹
to get help for APL primitives. Currently system functions and variables are not in )help, but I suppose extending file src/Help.def could easily add them.
Compiling APL is IMHO a wrong path. Too many problems, too little gain.
Best Regards, Jürgen Sauermann
On 10/16/19 5:01 PM, Rowan Cannaday wrote:
Thank you for the explanation Jürgen.
That makes intuitive sense. A shared-memory single threaded service is a reasonable abstraction.
Another approach, is to compile a subset of APL to an intermediate representation.
Is there a way to export the AST? in addition - is there an in-repl method of viewing help and/or arguments for system variables & functions?
By the way, a minor regression: segfaulting, but only after exiting. ``` )OFF ==================================================== SEGMENTATION FAULT thread: 0x7f8747766700 thread_cSegmentation fault ```
Thanks again, - Rowan
On Wed, Oct 16, 2019 at 12:06 PM Dr. Jürgen Sauermann <mail@jürgen-sauermann.de> wrote: Hi Blake,
it is sort of working, but I could well use some help in troubleshooting the remaining problems. I can help fixing them, but finding their root cause (and making them reproducible) is a different story.
My current interpretation of various benchmarks that Elias Mårtenson and myself did some years ago is that the bandwidth of the memory interface between the CPUs (or cores) and the memory is the limiting factor, and no matter how efficient the APL interpreter is, this bottleneck will dictate the speedup that can be achieved.
As an example, from 1985 to 1990, myself and 4 students had built a the hardware of a parallel APL machine with 32 CPUs and measured a speedup of close to 32 for sufficiently large vectors.
In contrast, if I remember correctly, then Elias achieved a speedup of 12 with 80 CPUs using the parallel feature of GNU APL. The only difference that I can see between our 1990 machine (called Datis-P-256 because the architecture could be scaled up to 256 processors) was the memory architecture:
Datis-P had one separate memory for each CPU, while current multicore boxes share their memory module(s) among different cores. That simply boils down to the fact that the memory bandwidth of Datis-P scaled with the number of processors, while the number of cores on a typical multi-core box does not. As long as this is the case, parallel APL remains severely limited in terms of the speedup that can be achieved.
Best Regards, Jürgen Sauermann
On 10/16/19 12:58 PM, Blake McBride wrote:
Greetings,
I think getting the parallel processing working is important. It may be that for various reasons the speedup in general cases is minimal and not worth the effort. However, I'd imagine that there are particular use-cases utilizing large arrays where the speedup would be substantial. That is when those types of enhancements would make APL a real benefit.
Thanks.
Blake
On Wed, Oct 16, 2019 at 5:27 AM Dr. Jürgen Sauermann <mail@jürgen-sauermann.de> wrote: Hi Rowan,
fixed in SVN 1191.
You should not be too enthusiastic, though, because the speed-ups that can be achieved are somewhat disappointing. And due to that, I haven't put too much effort into fixing faults (sometimes apl hangs on a semaphore when parallel execution is enabled).
Best Regards, Jürgen Sauermann
On 10/16/19 5:15 AM, Rowan Cannaday wrote:
Hello,
intrigued by the ability to parallelize APL, thought I'd try to test it:
`apl --cfg` followed by a line of '=' signs followed by `apl -q`:
configurable options: --------------------- ASSERT_LEVEL_WANTED=2 SECURITY_LEVEL_WANTED=0 (default) APSERVER_PATH=/tmp/GNU-APL/APserver (default) APSERVER_PORT=16366 (default) APSERVER_TRANSPORT=0 (default) CORE_COUNT_WANTED=2 DYNAMIC_LOG_WANTED=yes MAX_RANK_WANTED=8 (default) RATIONAL_NUMBERS_WANTED=yes SHORT_VALUE_LENGTH_WANTED=12, therefore: sizeof(Value) : 456 bytes sizeof(Cell) : 24 bytes sizeof(Value header): 168 bytes
VALUE_CHECK_WANTED=yes VALUE_HISTORY_WANTED=yes VF_TRACING_WANTED=no (default) VISIBLE_MARKERS_WANTED=yes
how ./configure was (probably) called: -------------------------------------- ./configure 'CORE_COUNT_WANTED=2' 'DEVELOP_WANTED=yes' 'VALUE_HISTORY_WANTED=yes' 'VISIBLE_MARKERS_WANTED=yes' '--enable-maintainer-mode'
BUILDTAG: --------- Project: GNU APL Version / SVN: 1.8 / 1190M Build Date: 2019-10-16 02:45:24 UTC Build OS: Linux 5.2.0-3-amd64 x86_64 config.status: 'CORE_COUNT_WANTED=2' 'DEVELOP_WANTED=yes' 'VALUE_HISTORY_WANTED=yes' 'VISIBLE_MARKERS_WANTED=yes' '--enable-maintainer-mode' Archive SVN: 1161
================================================================================
$ apl -q
==================================================== SEGMENTATION FAULT thread: 0x7f6078042e00 thread_contexts_count: 2 busy_worker_count: 0 active_core_count: 1 thread # 0: 0 RUN job: 0 no-name thread #-1: 0 RUN job: 0 no-name
---------------------------------------- -- Stack trace at main.cc:88 ---------------------------------------- 0x7F6078FD1BBB __libc_start_main 0x5631406C386D main 0x5631406CAD8D init_apl(int, char const**) 0x5631407E881B Parallel::init(bool) 0x563140832E2D Thread_context::init_parallel(CoreCount, bool) 0x7F60794E5B18 sem_init 0x7F60794E8510 0x5631406CA95A ======================================== ====================================================
|