1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
|
% language=us
\environment evenmore-style
% This appeared in TB 126, thanks to Karl Berry for fixing the English.
\startcomponent evenmore-pi
\startchapter[title=\TEX\ and Pi]
This is a short status report \footnote {This chapter appeared in \TUGBOAT\ 40:3.
Thanks to Karl Berry for corrections.} on Pi, not the famous version number of
\TEX\ (among other things), but the small machine, meant for education but
nowadays also used for Internet Of Things projects, process control and toy
projects. While the majority of \TEX\ installations run on an Intel processor,
the Raspberry Pi has an \ARM\ central processing unit. In fact, its main chip has
the same foundation as those found in settop boxes all around the world. It's
made for entertainment, not for number crunching.
At the \CONTEXT\ meetings, it has become tradition to play with electronic
gadgets. Every year we are curious what Harald K\"onig might bring this time. The
last couple of meetings we also had talks about using \TEX\ and \METAPOST\ for
designing (home|-|scale, automated) railroad systems, using \LUATEX\ for running
domotica applications, using \METAPOST\ for rendering high quality graphics from
data from appliances, presenting \TEX\ at computer and electronics bootcamps, and
more. This year Frans Goddijn also brought back memories of low speed modem
sounds, from the early days of \TEX\ support. It is these things that make the
meetings fun.
This year the meeting was in Belgium, close to the border of the Netherlands, and
on the way there Mojca Miklavec traveled via my home, where the contextgarden
compile farm runs on a server with plenty of cores, lots of memory and big disks.
But the farm also has an old Mac connected as well as a tiny underpowered
Raspberry Pi 2 for \ARM\ binaries that we had to fix: the small micro \SSD\ card
in it had finally given up. This is no surprise if you realize that it does a
daily compilation of the whole \TEX\ Live setup and also compiles \LUATEX,
\LUAMETATEX\ and pplib when changes occur. Replacing the card worked out but
nevertheless we decided to take the small machine with us to the meeting. We also
took an external (2.5 inch) \SSD\ box with us. The idea was to order a Raspberry
Pi~4 on location, the much praised successor of the older models, the one with
4~\GB\ of memory, real \USB~3 ports and proper Ethernet.
At the meeting Harald showed us that he had version~1, 3 and~4 machines with him
because he was looking into an energy control setup based on Zigbee devices. So
we had the full range of Pi's there to play with.
This is a long introduction but the message is that we are dealing with a small
but popular device with up to now four generations, using an architecture
supported in \TEX\ distributions. So how does that relate to \CONTEXT ? One of
the reasons for \LUAMETATEX\ going lean and mean is that computers are no longer
getting much faster and \quote {multiple small} energy|-|wise has more appeal
than \quote {one large}. So then the question is: how can we make \TEX\ run fast
on small instead of gambling on big becoming even bigger (which does not seem to
be happening anyway).
At the meeting Harald gave a talk \quotation {Which Raspberry Pi is the best for
\CONTEXT ?} and I will use his data to give an overview: see Table~\ref{rpispec}.
\starttabulate[||c|c|c|c|]
\BC model \BC 1 \BC 2 \BC 3 \BC 4 \NC \NR
\BC chipset \NC BCM2835 \NC BCM2835 \NC BCM2835 \NC BCM2835 \NC \NR
\BC CPU core \NC v6l rev 7 \NC v7l rev 5 \NC v7l rev 4 \NC v7l rev 3 \NC \NR
\BC cores \NC 1 \NC 4 \NC 4 \NC 4 \NC \NR
\BC free mem \NC 443080 \NC 948308 \NC 948304 \NC 3999784 \NC \NR
\BC idlemips \NC 997.08 \NC 38.40 \NC 38.40 \NC 108.00 \NC \NR
\BC bogomips \NC 997.08 \NC 57.60 \NC 76.80 \NC 270.00 \NC \NR
\BC read SD \NC 23.0 MB/s \NC 23.2 MB/s \NC 23.2 MB/s \NC 45.1 MB/s \NC \NR
\BC read USB \NC \NC 30.0 MB/s \NC 30.0 MB/s \NC 320.0 MB/s \NC \NR
\stoptabulate
After some discussion at the presentation we decided to discard the (absurd)
bogomips value for the tiny Pi~1 computing board and not take the values for the
others too seriously. But it will be clear that, especially when we consider the
external drive that things have improved. The table doesn't mention Ethernet
speed but because the~4 now has real support for it (instead of sharing the \USB\
bus) we get close to 1~\GB/s there.
The real performance test is of course processing a \TEX\ document and what
better to test than the \TEX\ book. The processing time in seconds, after initial
caching of files and fonts is:
\starttabulate[||c|c|c|c|]
\BC model \BC 1 \BC 2 \BC 3 \BC 4 \NC \NR
\BC \TEX book \NC 13.649 \NC 7.023 \NC 4.553 \NC 1.694 \NC \NR
\BC {\tt context --make} \NC \NC 19.949 \NC 11.796 \NC 6.034 \NC \NR
\BC {\tt context --make} TL \NC 89.454 \NC 46.578 \NC 29.256 \NC 14.146 \NC \NR
\stoptabulate
The test of making the \CONTEXT\ format using \LUATEX\ gives an indication of how
well the \IO\ performs: it loads the file database, some 460 \LUA\ modules and
355 \TEX\ source files. On my laptop with Intel i7-3840QM with 16GB memory and
decent \SSD\ it takes 3.5 seconds (and 1 second less for \LUAMETATEX\ because
there we don't compress the format file). Somehow a regular \TEXLIVE\
installation performs much worse than the one from the contextgarden.
We didn't test real \CONTEXT\ documents at the meeting but when I came home the
Pi~4 was bound again to the compile farm. Harald and Mojca had prepared the
machine to boot from the internal micro \SSD\ and use the external disk for the
rest. So, when we could compile \LUAMETATEX\ again, I made an \ARM\ installer for
\LMTX, and after that could not resist doing a simple test. First of course came
generating the format. It took 6.3 seconds to make one, which is a bit more than
Harald measured. I see a hiccup at the end so I guess that it has to do with the
(external) disk or maybe there is some throttling going on because the machine
sits on top of a (warm) server.
More interesting was testing a real document: the upcoming \LUAMETATEX\ manual.
It has 226 pages, uses 21 font files, processes 225 \METAPOST\ graphics, and in
order to get it \LUAMETATEX\ does more than 50\% of the work in \LUA, including
all font and backend|-|related operations. On my laptop it needs 9.5~seconds and
on the Pi~4 it uses 33~seconds. Of course, if I take a more modern machine than
this 8|-|year|-|old workhorse, I probably need half the time, but still the
performance of the Raspberry Pi~4 is quite impressive. It uses hardly any energy
and can probably compete rather well with a virtual machine on a heavily loaded
machine. It means that when we ever have to upgrade the server, I can consider
replacement by an Ethernet switch, with power over Ethernet, connected to a bunch
of small Raspberries, also because normally one would connect to some shared
storage medium.
Because I was curious how the dedicated small Fitlet that I use for controlling
my lights and heating performs I also processed the manual there. After making
the format, which takes 6~seconds, processing the manual took a little less than
30~seconds. In that respect it performs the same as a Raspberry Pi~4. But, inside
that small (way more expensive) computer is an dual core AMD A10 Micro-6700T APU
(with AMD Radeon R6 Graphics), running a recent 64|-|bit Ubuntu. It does some
2400 bogomips (compare that to the values of the Pi). I was a bit surprised that
it didn't outperform the Raspberry because the (fast \SSD) disk is connected to
the main board and it has more memory and horsepower. It might be that in the end
an \ARM\ processor is simply better suited for the kind of byte juggling that
\TEX\ does, where special \CPU\ features and multiple cores don't contribute
much. It definitely demonstrates that we cannot neglect this platform.
\stopchapter
\stopcomponent
|
