AWK Reloaded

Last year I compared the performance of 3 AWK interpreters, NAWK, GAWK and MAWK. For the test I used 3 of my .awk scripts (available under Beerware). But the data I processed with them was confidential. Any way, NAWK won 1/3, MAWK 2/3 (with astonishing leads), GAWK was the clear looser with abysmal performance in 2/3 tests.

Recently I developed a run time interpreter for the Heidenhain NC (Numerical Control) language. The code of the script as well as the program it interprets are confidential, unfortunately. But the results are interesting nonetheless.

Tests

Test Environment

Unfortunately I cannot run the tests on the same machine as last time, it was stolen during a visit in the UK last winter.

So this time the tests are run on its replacement, an Intel Haswell Core i-7 (2 cores, 4 pipelines) at 2.4 GHz under FreeBSD 10 r267867 (amd64).

AWK Versions

• nawk 20121220 (FreeBSD)
• gawk 4.1.1
• mawk 1.3.4

hhrti.awk [1 pt/s]

This is a test run with the aforementioned run time interpreter. There is a more in depth explanation at the end of this article.

150.43 s

149.41 s

149.29 s

67.40 s

67.86 s

67.61 s

48.97 s

47.48 s

48.02 s

Memory usage (maximum resident set size):

2864 k

4240 k

2760 k

xml.awk [100 pt/s]

This is one of the tests run last time, to confirm that the interpreters still compare similarly with the previous scripts, despite the updated test platform. That seems to be the case here.

0.16 s

0.16 s

0.16 s

0.57 s

0.57 s

0.58 s

0.02 s

0.02 s

0.02 s

Conclusions

Consistently with the previous performance tests, MAWK takes the lead. What is surprising that GAWK performs well with only 1.38 times the run time of MAWK, which is a far cry from the abysmal performance it exhibited in some of the other tests. A quick rerun of the previous tests shows the same performance gaps as before, so neither the slight version changes nor the new compiler version (clang 3.4.1) introduced a performance boost in GAWK.

The real surprise is the performance of NAWK. This is the first test case where it performs worse than GAWK, with a runtime factor of 3.0. That's a far cry from GAWK's sad >25 in the xml.awk case, but still it hints to a bottleneck in NAWK.

Differences to Previous Tests

This test is a lot less array heavy than the dbc2c.awk and xml.awk test cases. The parsing stage barely takes any time after that only small local arrays are used for temporary tokenizing. The most time consuming operation seems to be evaluating arithmetic, because whenever an operation is performed it creates a number of copy operations. Depending on the operator all following tokens need to be shifted one or two places.

Bottleneck Test

In order to verify the assumption that copies in arrays might be responsible I created a small script that performs this operation repeatedly:

BEGIN {
 TXT = "111 222 333 444 555 666 777 888 999 000"
 REPEAT = 100000
 if (ARGC > 1) {
  REPEAT = ARGV[--ARGC]
  delete ARGV[ARGC]
 }
 srand() # Seed

 # Perform the test this many times
 for (i = 1; i <= REPEAT; i++) {

  # Create an array with tokens
  len = split(TXT, a)
  a[0] = len # Store the length in index 0, this is very
             # convenient in real apps with lots of arrays

  # Test case, delete a random field until none
  # are left
  while (a[0]) {
   # Select a random entry to delete
   del = int(rand() * 65536) % a[0] + 1

   # Shift the following tokens left
   for (p = del; p < a[0]; p++) {
    a[p] = a[p + 1]
   }
   # Delete the tail
   delete a[a[0]--]
  }
 }
}

bottleneck.awk [10 pt/s]

This artificial test seems to confirm this, by reproducing the same performance pattern and amplifying the performance problem of NAWK. The script was run with 200000 repetitions.

12.24 s

12.28 s

12.17 s

2.29 s

2.29 s

2.28 s

1.68 s

1.63 s

1.69 s

Memory usage (maximum resident set size):

2488 k

3596 k

2476 k

---

The Heidenhain Real Time Interpreter

The Heidenhain NC language can be used to control an NC mill. I.e. access various functions of the machine, such as cooling systems, automatic tool changers and provide milling instructions. Additionally it has programming instructions that can be used to make on the fly calculations and decisions. The purpose of the interpreter is to perform arithmetic and conditional flow in advance.

The need for this arose with a program written for a research project, which is so computation heavy that it causes the machine to stutter.

The interpreter works in two stages, a code parsing stage and an evaluation stage.

Parsing

In this stage every command is stored in a one-way linked list. Additional code files may be called within a program, those are parsed after the current file has been completed and appended to the same list.

The Heidenhain NC language has several kinds of commands, most of these are pretty static, they access machine functions, or describe target coordinates or curves. These kinds of commands are what the interpreter outputs in the evaluation stage.

The other kind of commands provide arithmetic and program flow:

• Variable assignments
• Arithmetic expressions (really part of variable assignments)
• Labels
• Label calls
• Conditional label calls (i.e. IF)
• Program calls

The list is not complete, but it should get the idea across.

Every list entry is classified during parsing stage, and some are preprocessed. E.g. labels and subprogram entries are recorded in an associative array so they can be branched to in the evaluation stage.

Evaluation

In this stage the interpretation is performed. The program starts with an empty call stack at the first parsed code line. Each line is evaluated according to its classification.

• Variable assignments are evaluated and stored in an associative array containing name, value pairs
• Label calls are performed
• Conditions are evaluated and either branch to a label, that is fetched from the array recorded during parsing, or continue with the next line
• Calls to other programs cause a reference to the current code line to be pushed on the stack
• Ends of programs cause a call back to the code line recorded on the stack, if the stack is empty the interpreter terminates
• …

Every command that is not classified for special treatment receives the following default treatment:

1. Substitute variables with their current values
2. Output the command

The result is a flat NC program that does no longer contain arithmetic and conditional code. E.g:

0    BEGIN PGM mandelbrot_kante MM
1    BLK FORM 0.1 Z X+0 Y-90.0000 Z-50
2    BLK FORM 0.2 X+220.0000 Y+90.0000 Z+0
3    TOOL CALL 4 Z S32000 F8000
4    M3
5    L Z+20 FMAX
6    L X+110.0000 Y-52.2387 FMAX
7    L Z+2 FMAX
8    L Z-0.5000 F800
9    L X+110.2000 Y-52.2387 F8000
[...]
7758 L X+109.4000 Y-52.8387 F8000
7759 L X+109.4000 Y-52.6387 F8000
7760 L X+109.4000 Y-52.4387 F8000
7761 L X+109.4000 Y-52.2387 F8000
7762 L X+109.6000 Y-52.2387 F8000
7763 L X+109.8000 Y-52.2387 F8000
7764 L X+110.0000 Y-52.2387 F8000
7765 L X+110.2000 Y-52.2387 F8000
7766 L Z+50 FMAX
7767 END PGM mandelbrot_kante MM