Re: Absolute limits of rank 2 bool matrix size in GNU APL?

[Elias Mårtenson]

Actually, the default memory allocator in glibc uses mmap and not sbrk, if I'm not mistaken. However, they both do the same thing at the end of the day, and it's to request more pages from the kernel.

Unfortunately, Linux malloc never returns NULL. Even if you try to allocate a petabyte in one allocation. You can write to this memory and new pages will be created as you write, and at some point your write will fail with a SEGV because there are no free page left.

What's even worse is that once you start to fail, the kernel will randomly kill processes until it gets more free pages. If that sounds idiotic, that's because it is, but that's how it works. 

You can configure the kernel to use not do this, but unfortunately you're going to have other problems if you do, primarily because all Linux software is written with the default behaviour in mind.

You can try it right now. Write a C program that allocates a few TB of RAM and see what happens. Actually, nothing will happen until you start writing to this memory.

Den ons 29 dec. 2021 03:09Blake McBride <blake1024@gmail.com> skrev:

Hi Jürgen,

First, to be sure you know where I'm coming from.  I know malloc() and sbrk() pretty well.  I've written malloc's before.

It is utter news to me that malloc would return a pointer that you can't use - even in over-commit situations.  Theoretically, in over-commit situations, if you access a pointer that's paged out, the system should page it in first.  If that doesn't happen, that's a bug in the OS.  I'd create a simple program that mimics this behavior and submit it as a bug report.  (What I almost always find is that during the process of creating the sample program, I find the actual bug in my own program.)

What I'd try to do is write my own malloc that does the following:

1.  Checks for the amount of real memory available using something I took from the free command.

2.  If there is enough memory, call the real malloc and return it.

3.  If there is not enough real memory, rather than calling malloc and causing paging you can just return a null.

Of course, I understand that sbrk() adjusts the system breakpoint essentially giving you more heap space.  Malloc manages that free space that's already been taken from the OS via sbrk().

With regard to item #1, you'd have to have an intelligent way of determining when to calculate the true available space.  You can't do it for 16 bytes, and you don't want to do it for every allocation.  On the other hand for large allocations or if you haven't queried the system for some time, it makes sense to update your opinion of the available space.

Thanks.

Blake

On Tue, Dec 28, 2021 at 12:50 PM Dr. Jürgen Sauermann <mail@jürgen-sauermann.de> wrote:

Hi Blake,

I know. The problem is that other processes may, just as GNU APL does, allocate memory
successfully via malloc(). malloc, on "success" returns a non-zero virtual memory address.

The fact that malloc returns a non-zero address (= indicating malloc() success) does unfortunately
not guarantee that you would be able to access that memory. You may or may
not get a page fault when accessing the memory. The point in time when you call 'free' or
similar tools (which basically read /proc/memory tell you the situation at that time which can
be very different from the situation when you fill your virtual memory with real data (which
assigns a physical memory page to the successfully allocated virtual memory page) and
that may fail. Even worse, when this happen then your machine has no more memory and
the C++ exception handlers will also fails due to lack of memory and the process is terminated
without any chance to raise a WS full or even remain in apl.

Best Regards,
Jürgen

On 12/28/21 6:52 PM, Blake McBride wrote:

Since the Linux "free" command knows the difference between how much real memory vs. virtual memory is available, it may be useful to use a similar method.

Blake

On Tue, Dec 28, 2021 at 11:34 AM Dr. Jürgen Sauermann <mail@jürgen-sauermann.de> wrote:

Hi Russ,

it has turned out to be very difficult to find a reliable way to figure the memory that is available for
a process like the GNU APL interpreter. One (of several) reasons for that is that GNU linux tells you
that it has more memory than it really has (so-called over-commitment). You can malloc() more memory
than you have and malloc will return a virtual memory (and no error) when you ask for it. However, if you
access the virtual memory at a later point in time (i.e. requesting real memory for it) then the process
may crash badly if that real memory is not available.

GNU APL deals with this by assuming (by default) that the total memory available for the GNU APL process is about 2 GB,

You can increase that amount (outside apl) be setting a larger memory limit, probably with:

ulimit --virtual-memory-size or ulimit -v (depending on platform).

in the shell or script before starting apl. However, note that:

* ulimit -v ulimited will not work because this is the default and GNU APL will apply its default of 2 GB in this case,
* the WS FULL behaviour of GNU APL (ans ⎕WA) will become unreliable. You must ensure that GNU APL will get
  as much memory as you promised it with ulimit, and
* all GNU APL cells have the same size (e.g. 24 byte on a 64-bit CPU, see apl -l37) even if the cells contain only
 Booleans.

The workaround for really large Boolean arrays is to pack them into the 64-bits of a GNU APL integer
and maybe use the bitwise functions (⊤∧, ⊤∨, ...) of GNU APL to access them group-wise.

Best Regards,
Jürgen

On 12/28/21 3:53 AM, Russtopia wrote:

Hi, doing some experiments in learning APL I was writing a word frequency count program that takes in a document, identifies unique words and then outputs the top 'N' occurring words.

The most straightforward solution, to me, seems to be ∘.≡ which works up to a certain dataset size. The main limiting statement in my program is

wordcounts←+⌿ (wl ∘.≡ uniqwords)

.. which generates a large boolean array which is then tallied up for each unique word.

I seem to run into a limit in GNU APL. I do not see an obvious ⎕SYL parameter to increase the limit and could not find any obvious reference in the docs either. What are the absolute max rows/columns of a matrix, and can the limit be increased? Are they separate or a combined limit?

      5 wcOuterProd 'corpus/135-0-5000.txt'    ⍝⍝ 5000-line document
Time: 26419 ms
  the   of   a and  to
 2646 1348 978 879 858
      ⍴wl
36564
      ⍴ uniqwords
5695

      5 wcOuterProd 'corpus/135-0-7500.txt'   ⍝⍝ 7500-line document
WS FULL+
wcOuterProd[8]  wordcounts←+⌿(wl∘.≡uniqwords)
                              ^             ^
      ⍴ wl
58666
      ⍴ uniqwords
7711

I have an iterative solution which doesn't use a boolean matrix to count the words, rather looping through using pick/take and so can handle much larger documents, but it takes roughy 2x the execution time.

Relating to this, does GNU APL optimize boolean arrays to minimize storage (ie., using larger bit vectors rather than entire ints per bool) and is there any clever technique other experience APLers could suggest to maintain the elegant 'loop-free' style of computing but avoid generating such large bool matrices? I thought of perhaps a hybrid approach where I iterate through portions of the data and do partial ∘.≡ passes but of course that complicates the algorithm.

[my 'outer product' and 'iterative' versions of the code are below]

Thanks,

-Russ

---

#!/usr/local/bin/apl --script
 ⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝
⍝                                                                    ⍝
⍝ wordcount.apl                        2021-12-26  20:07:07 (GMT-8)  ⍝
⍝                                                                    ⍝
 ⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝⍝

⍝ function edif has ufun1 pointer 0!

∇r ← timeMs; t
  t ← ⎕TS
  r ← (((t[3]×86400)+(t[4]×3600)+(t[5]×60)+(t[6]))×1000)+t[7]
∇

∇r ← lowerAndStrip s;stripped;mixedCase
 stripped ← '               abcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyz*'
 mixedCase ← ⎕av[11],' ,.?!;:"''()[]-ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz'
 r ← stripped[mixedCase ⍳ s]
∇

∇c ← h wcIterative fname
  ⍝⍝;D;WL;idx;len;word;wc;wcl;idx
  ⍝⍝ Return ⍒-sorted count of unique words in string vector D, ignoring case and punctuation
  ⍝⍝ @param h(⍺) - how many top word counts to return
  ⍝⍝ @param D(⍵) - vector of words
  ⍝⍝⍝⍝
  D ← lowerAndStrip (⎕fio['read_file'] fname)  ⍝ raw text with newlines
  timeStart ← timeMs
  D ← (~ D ∊ ' ') ⊂ D ⍝ make into a vector of words
  WL ← ∪D
  ⍝⍝#DEBUG# ⎕ ← 'unique words:',WL
  wcl ← 0⍴0
  idx ← 1
  len ← ⍴WL
count:
  ⍝⍝#DEBUG# ⎕ ← idx
  →(idx>len)/done
  word ← ⊂idx⊃WL
  ⍝⍝#DEBUG# ⎕ ← word
  wc ← +/(word≡¨D)
  wcl ← wcl,wc
  ⍝⍝#DEBUG# ⎕ ← wcl
  idx ← 1+idx
  → count
done:
  c ← h↑[2] (WL)[⍒wcl],[0.5]wcl[⍒wcl]
  timeEnd ← timeMs
  ⎕ ← 'Time:',(timeEnd-timeStart),'ms'
∇

∇r ← n wcOuterProd fname
  ⍝⍝ ;D;wl;uniqwords;wordcounts;sortOrder
  D ← lowerAndStrip (⎕fio['read_file'] fname)  ⍝ raw text with newlines
  timeStart ← timeMs
  wl ← (~ D ∊ ' ') ⊂ D
  ⍝⍝#DEBUG# ⎕ ← '⍴ wl:', ⍴ wl
  uniqwords ← ∪wl
  ⍝⍝#DEBUG# ⎕ ← '⍴ uniqwords:', ⍴ uniqwords
  wordcounts ← +⌿(wl ∘.≡ uniqwords)
  sortOrder ← ⍒wordcounts
  r ← n↑[2] uniqwords[sortOrder],[0.5]wordcounts[sortOrder]
  timeEnd ← timeMs
  ⎕ ← 'Time:',(timeEnd-timeStart),'ms'
∇