Filippo Mantovani: ARM for HPC

On 23-Oct-2017 Filippo Mantovani held a talk in Darmstadt on “Mobile technology for production-ready high-performance computing systems: The path of the Mont-Blanc project”. Unfortunately I was unable to attend, but Mr. Mantovani sent me his Darmstadt Seminar slides. As his slides and documents are very interesting to people using or intending to use ARM in HPC, I copy these documents here, so they are easily available. I also copied a report on “MB3_D6.4 Report on application tuning and optimization on ARM platform“.

Advertisements

Optimal Product Portfolio

1. Prerequisites

Assume one owns n>0 different products, n\in\mathbb{N}, usually n\approx 100. Each product has a cost c_i>0, c_i\in\mathbb{R}, associated with this ownership, i=1,\ldots, n. Let N=\left\{1,\ldots,n\right\}, and \mathbb{P}(N) be the powerset of N, i.e., the set of all subsets. The powerset \mathbb{P}(N) has 2^n elements.

There are two real-valued functions f and g operating on \mathbb{P}(N), i.e., f,g\colon\mathbb{P}(N)\to\mathbb{R^+}. Function g(I) denotes the sum of the cost of the product set given by I\in\mathbb{P}(N), g is therefore easy to evaluate. I.e.,

\displaystyle{      g(I) = \sum_{i\in I} c_i  }

Function f(I) denotes the buying cost if one can dispose the set of products given by I\in\mathbb{P}(N). This means, it costs money to get rid of the products given by set I. As the products have interrelationships, f is more costly to evaluate than g and is given by some table lookup. Function f does not depend on c_i.

Some notes on f: Assume a non-negative matrix A=(a_{ij}), where a_{ij} denotes the cost to decouple the i-th product from the j-th. Then

\displaystyle{      f(I) = \sum_{i\in I} \sum_{j\in N} a_{ij}  }

Usually a_{ij}=0, if i<j, i.e., A is upper triangular, because decoupling product i from product j does not need to be done again for decoupling the other direction from j to i. More zeros in above sum are necessary if decoupling product i is sufficient if done once and once only. In practical application cases f depends on a real-valued parameter \gamma\in[0,1] giving different solution scenarios.

Example: For three products let c_1=8, c_2=4, c_3=5, thus cost of ownership is g(\emptyset)=0, g(\left\{1\right\})=8, g(\left\{1,2\right\})=8+4=12, g(\left\{1,3\right\})=8+5=13, g(\left\{1,2,3\right\})=8+4+5=17, and so on for the rest of 2^3-5=3 sets. Function f gives positive values in a similar fashion, although, as mentioned before, these values are not related to c_i.

2. Problem statement

Find the set I of products which gives the least cost, i.e., find I for

\displaystyle{      \min_{I\in\mathbb{P}(N)}\left(f(I)-g(I)\right)  }

Rationale: One invests f(I) to get rid of the products in set I but saves ownership costs given by g(I).

I does not need to be unique. If f depends on \gamma then I will likely also depend on \gamma.

3. Challenges

As n can be “large”, evaluating all possible combinations becomes prohibitive on todays computers. Maybe one of our gentle readers knows a method to find the optimum, or maybe just a nearby solution to above problem, or a probabilistic approach. Any comment is welcome.

Statistics of this Blog: Crossed 60.000 Views

This blog was viewed more than 60.000 times since its inception and had more than 45.000 visitors. I wrote about the development of this blog as follows:

  1. 2014/05/07: 5,000 Views
  2. 2014/09/10: 10,000 Views
  3. 2014/12/27: 15,000 Views
  4. 2015/04/23: 20,000 Views
  5. 2016/01/24: 30,000 Views
  6. 2016/11/12: 40,000 Views
  7. 2017/05/18: 50,000 views

The averages per month are:

Between 1,000 and 2,000 views per month in bar chart form:

As Countries, USA, Germany, and India at the top:

Referrers, Google by far the most important referrer:

The most popular blog posts are:

C Pointer Surprises

An article by Krister Walfridsson on C pointers are not hardware pointers demonstrated that even adjacent integer variables having the same hardware address may compare unequal regarding C pointers.

See the following C program:

#include <stdio.h>

int main(int argc, char *argv[]) {
        int x, y;
        int *p = &x + 1;
        int *q = &y;
        printf("%p %p %d\n", (void*)p, (void*)q, p == q);
        return 0;
}

You have to compile with optimization enabled, e.g., cc -O3. Otherwise gcc adds some stuff between variables. On AMD/Intel/ARM CPUs the output looks something like this:

0xbe849afc 0xbe849afc 0

I.e., the pointers point to the same address, but the pointer comparison gives “false”.

Added 06-Aug-2017: As hinted by the comment given by Ashwin Nanjappa below, the compiler actually does not generate compare instructions, but rather just adds 0=false.

Disassembling

$ cc -Wall -O3 -c ptrcomp.c
$ objdump -d ptrcomp.o

gives

ptrcomp.o:     file format elf64-x86-64

Disassembly of section .text.startup:

0000000000000000 <main>:
   0:   48 83 ec 18             sub    $0x18,%rsp
   4:   48 8d 3d 00 00 00 00    lea    0x0(%rip),%rdi        # b <main+0xb>
   b:   31 c9                   xor    %ecx,%ecx
   d:   48 8d 54 24 04          lea    0x4(%rsp),%rdx
  12:   64 48 8b 04 25 28 00    mov    %fs:0x28,%rax
  19:   00 00
  1b:   48 89 44 24 08          mov    %rax,0x8(%rsp)
  20:   31 c0                   xor    %eax,%eax
  22:   48 89 d6                mov    %rdx,%rsi
  25:   e8 00 00 00 00          callq  2a <main+0x2a>
  2a:   48 8b 4c 24 08          mov    0x8(%rsp),%rcx
  2f:   64 48 33 0c 25 28 00    xor    %fs:0x28,%rcx
  36:   00 00
  38:   75 07                   jne    41 <main+0x41>
  3a:   31 c0                   xor    %eax,%eax
  3c:   48 83 c4 18             add    $0x18,%rsp
  40:   c3                      retq
  41:   e8 00 00 00 00          callq  46 <main+0x46>

Xoring oneself gives zero.

Text Analysis using Concordance

When analyzing longer text, especially if this text was written by oneself, it helps to read the text in a different way, here using a concordance.

Assume your text is provided as PDF. Convert PDF to text using pdftotext, which part of package poppler. Replace line breaks in text file with spaces using below C program (called linebreak.c):

#include <stdio.h>

int main(int argc, char *argv[]) {
        int c, flag=0;
        FILE *fp;

        if (argc >= 2) {
                if ((fp = fopen(argv[1],"rb")) == NULL)
                        return 1;
        } else {
                fp = stdin;
        }

        while ((c = fgetc(fp)) != EOF) {
                if (c == '\n') {
                        flag += 1;
                        if (flag > 1) { putchar(c); flag = 0; }
                        else putchar(' ');
                } else {
                        flag = 0;
                        putchar(c);
                }
        }

        return 0;
}

Then generate a list of (single) words with below Perl program:

#!/bin/perl -W
# Print word concordances

use strict;

my (%H,@F);

while (<>) {
        chomp;
        s/\s+$//;       # rtrim
        @F = split;
        foreach my $w (@F) {
                $w =~ s/^\s+//; # ltrim
                $w =~ s/\s+$//; # rtrim
                $H{$w} += 1;
        }
}

foreach my $w (sort keys %H) {
        printf("\t%6d\t%s\n",$H{$w},$w);
}

To print all word pairs replace above loop with

while (<>) {
        chomp;
        s/\s+$//;       # rtrim
        @F = split;
        for(my $i=0; $i<$#F; ++$i) {
                $F[$i] =~ s/^\s+//;     # ltrim
                $F[$i] =~ s/\s+$//;     # rtrim
                $F[$i+1] =~ s/^\s+//;   # ltrim
                $F[$i+1] =~ s/\s+$//;   # rtrim
                $H{$F[$i] . " " . $F[$i+1]} += 1;
        }
}

Similar, for word triples replace the loop with

while (<>) {
        chomp;
        s/\s+$//;       # rtrim
        @F = split;
        for(my $i=0; $i+1<$#F; ++$i) {
                $F[$i] =~ s/^\s+//;     # ltrim
                $F[$i] =~ s/\s+$//;     # rtrim
                $F[$i+1] =~ s/^\s+//;   # ltrim
                $F[$i+1] =~ s/\s+$//;   # rtrim
                $F[$i+2] =~ s/^\s+//;   # ltrim
                $F[$i+2] =~ s/\s+$//;   # rtrim
                $H{$F[$i] . " " . $F[$i+1] . " " . $F[$i+2]} += 1;
        }
}

Printing concordances using Perl hashes is very simple, as one can see.

Here is an example from the man-page of expect using below sequence of commands:

( TERM=dumb; man expect ) | linebreak | word3concord | sort -r

Truncated result is

            16  For example, the
            13  example, the following
            12  the current process.
             9  the end of
             8  using Expectk, this
             8  this option is
             8  sent to the
             8  flag causes the
             8  body is executed
             8  Expectk, this option
             8  (When using Expectk,
             7  to the current
             7  the spawn id
             7  the most recent
             7  the current process
             7  the corresponding body
             7  option is specified
             7  is specified as
             7  corresponding body is
             7  by Don Libes,
             7  be used to
             6  set for the
             6  of the current
             6  is set for
             6  is an alias

Contributing to Hugo Static

The discussion forum for Hugo contains a description: Hugo development – how to contribute code. Also see Contributing to Hugo.

1. Preparation

First set GOPATH as

export GOPATH=$HOME/tmp/H

then

cd $GOPATH

Fetch source with go get

time go get -u -v github.com/spf13/hugo

takes around 1-2 minutes as it has to download almost 200MB.

Now change to the Hugo source code and compile

cd src/github.com/spf13/hugo/
time make hugo

Compilation from scratch takes roughly 1-2 minutes. Recompiling a single file usually takes less than 10 seconds.

In the same directory, run test-cases with

time make check

which takes less than a minute.

All timings are on an AMD FX(tm)-8120 Eight-Core Processor clocked with 3.1 GHz running Linux 4.11.3, and using Go 1.8.3.

2. Fork in Github, git branch and pull-request

Fork https://github.com/spf13/hugo by pressing the “Fork” icon:

Move original Git repository out of your way, clone the new fork, add or modify files as required, add, and commit them:

cd $GOPATH/src/github.com/spf13/
mv hugo hugo.original
time git clone git@github.com:eklausme/hugo.git

cd hugo
git branch YOURNAME
git checkout YOURNAME

go fmt
git add YOURFILE
git commit

A git clone of hugo alone takes less than 10 seconds. Watch out to run go fmt before git add.

Contributors are asked to provide single commits. In case you have multiple, then squash them into one, i.e., git rebase -i and git push -f.

Finally press the pull-request button in Github:

Be prepared to wait weeks or even months before your pull-request will be accepted or even rejected, so patience is required.

HP ePrint Obsolescense

HP (Hewlett Packard) manufacturer of printers of various sorts (laser, ink) unfortunately again decided to annoy its loyal customers. Repeatedly HP updated firmware in printers so that do not work with alternate ink, see for example Disabling 3rd-party ink ensures “best printing experience”. Now for the last years they disabled the so called apps (ePrint functionality): after a few years the printer can no longer connect to the web-services of HP. See, for example, HP Apps Service Retired on several printers.

In my case, I bought a CM1415 in November 2011. It can no longer connect to HP web-services. Therefore I can no longer send e-mails to the printers which then get printed. In November 2014 I bought an M276nw, a similar printer as the CM1415. This model can still connect to HP web-services. So it looks like after six years HP silently disables functionality. These web-services also offer other services, like weather forecasts, news, sudokus, etc.

These chicanery make it clear that customers should not trust cloud-services, or at least have a contingency plan if these services do not work, or have prices which are ridiculous. Recently the price increase by Firebase made headlines, see Firebase Costs Increased by 7,000%!