Updated: 23 July 2006
Table Of Contents


ppmtompeg - encode an MPEG-1 bitstream


ppmtompeg [options] parameter-file


This program is part of Netpbm.

ppmtompeg produces an MPEG-1 video stream. MPEG-1 is the first great video compression method, and is what is used in Video CDs (VCD). ppmtompeg originated in the year 1995. DVD uses a more advanced method, MPEG-2. There is an even newer method called MPEG-4 which is also called Divx. I don't know where one finds that used.

There's technically a difference between a compression method for video and an actual file (stream) format for a movie, and I don't know if it can be validly said that the format of the stream ppmtompeg produces is MPEG-1.

Mencoder from the Mplayer package is probably superior for most video format generation needs, if for no other reason than that it is more popular.

The programming library PM2V generates MPEG-2 streams.

Use Mplayer (not part of Netpbm) to do the reverse conversion: to create a series of PNM files from an MPEG stream.

param_file is a parameter file which includes a list of input files and other parameters. The file is described in detail below.

To understand this program, you need to understand something about the complex MPEG-1 format. One source of information about this standard format is the section Introduction to MPEG in the Compression FAQ.


The -gop, -combine_gops, -frames, and -combine_frames options are all mutually exclusive.

-stat stat_file
This option causes ppmtompeg to append the statistics that it write to Standard Output to the file stat_file as well. The statistics use the following abbreviations: bits per block (bpb), bits per frame (bpf), seconds per frame (spf), and bits per second (bps).

These statistics include how many I, P, and B frames there were, and information about compression and quality.

-quiet num_seconds
causes ppmtompeg not to report remaining time more often than every num_seconds seconds (unless the time estimate rises, which will happen near the beginning of the run). A negative value tells ppmtompeg not to report at all. 0 is the default (reports once after each frame). Note that the time remaining is an estimate and does not take into account time to read in frames.
causes ppmtompeg to run silently, with the only screen output being errors. Particularly useful when reading input from stdin.
This option prevents ppmtompeg from printing a summary line for each frame
forces ppmtompeg to use a more accurate, yet more computationally expensive version of the DCT.
-gop gop_num
causes ppmtompeg to encode only the numbered GOP (first GOP is 0). The parameter file is the same as for normal usage. The output file will be the normal output file with the suffix .gop.gop_num. ppmtompeg does not output any sequence information.
causes ppmtompeg simply to combine some GOP files into a single MPEG output stream. ppmtompeg inserts a sequence header and trailer. In this case, the parameter file needs only to contain the SIZE value, an output file, and perhaps a list of input GOP files (see below). If you don't supply a list of input GOP files is used, then ppmtompeg assumes you're using the same parameter file you used when you created the input (with the -gop option) and calculates the corresponding gop filenames itself. If this is not the case, you can specify input GOP files in the same manner as normal input files -- except instead of using INPUT_DIR, INPUT, and END_INPUT, use GOP_INPUT_DIR, GOP_INPUT, and GOP_END_INPUT. If no input GOP files are specified, then the default is to use the output file name with suffix .gop.gop_num, with gop_num starting from 0, as the input files.

Thus, unless you're mixing and matching GOP files from different sources, you can simply use the same parameter file for creating the GOP files (-gop) and for later turning them into an MPEG stream (-combine_gops).

-frames first_frame last_frame
This option causes ppmtompeg to encode only the frames numbered first_frame to last_frame, inclusive. The parameter file is the same as for normal usage. The output will be placed in separate files, one per frame, with the file names being the normal output file name with the suffix .frame.frame_num. No GOP header information is output. (Thus, the parameter file need not include the GOP_SIZE value)

Use ppmtompeg -combine_frames to combine these frames later into an MPEG stream.

This option causes ppmtompeg simply to combine some individual MPEG frames (such as you might have created with an earlier run of ppmtompeg -frames) into a single MPEG stream. Sequence and GOP headers are inserted appropriately. In this case, the parameter file needs to contain only the SIZE value, the GOP_SIZE value, an output file, and perhaps a list of frame files (see below).

The parameter file may specify input frame files in the same manner as normal input files -- except instead of using INPUT_DIR, INPUT, and END_INPUT, use FRAME_INPUT_DIR, FRAME_INPUT, and FRAME_END_INPUT. If no input frame files are specified, then the default is to use the output file name with suffix .frame.frame_num, with frame_num starting from 0, as the input files.

This option causes ppmtompeg to run any remote processes "nicely," i.e. at low priority. (This is relevant only if you are running ppmtompeg in parallel mode. Otherwise, there are no remote processes). See 'man nice.'
-max_machines num_machines
This option causes ppmtompeg to use no more than num_machines machines as slaves for use in parallel encoding.
This option causes ppmtompeg to include the signal-to-noise ratio in the reported statistics. Prints SNR (Y U V) and peak SNR (Y U V) for each frame. In summary, prints averages of luminance only (Y). SNR is defined as 10*log(variance of original/variance of error). Peak SNR is defined as 20*log(255/RMSE). Note that ppmtompeg runs a little slower when you use this option.
This option causes ppmtompeg to report the mean squared error per block. It also automatically reports the quality of the images, so there is no need to specify -snr then.
-bit_rate_info rate_file
This option makes ppmtompeg write bit rate information into the file rate_file. Bit rate information is bits per frame, and also bits per I-frame-to-I-frame.
This option causes ppmtompeg to print a histogram of the motion vectors as part of statistics. There are three histograms -- one for P frame, one for forward B frame, and one for backward B frame motion vectors.

The output is in the form of a matrix, each entry corresponding to one motion vector in the search window. The center of the matrix represents (0,0) motion vectors.

This option causes ppmtompeg to print to Standard Output messages that narrate the communication between the machines when you run ppmtompeg in parallel mode.
This option causes ppmtompeg to print to Standard Output messages that narrate the progress of the conversion on the various machines when you run ppmtompeg in parallel mode.


The parameter file must contain the following lines (except when using the -combine_gops or -combine_frames options):

PATTERN pattern
This statement specifies the pattern (sequence) of I frames, P frames, and B frames. pattern is just a sequence of the letters I, P, and B with nothing between. Example:

See I Frames, P Frames, B Frames.

OUTPUT output file
This names the file where the output MPEG stream goes.
INPUT_DIR directory
This statement tells where the input images (frames) come from. If each frame is in a separate file, directory is the directory where they all are. You may use . to refer to the current directory. A null directory refers to the root directory of the system file tree.

To have ppmtompeg read all the frames serially from Standard Input, specify

    INPUT_DIR stdin
This line must be followed by a list of the input files (in display order) and then the line END_INPUT.

There are three types of lines between INPUT and END_INPUT. First, a line may simply be the name of an input file. Second, the line may be of the form single_star_expr [x-y]. single_star_expr can have a single * in it. It is replaced by all the numbers between x and y inclusive. So, for example, the line tennis*.ppm [12-15] refers to the files tennis12.ppm, tennis13.ppm, tennis14.ppm, tennis15.ppm.

Uniform zero-padding occurs, as well. For example, the line football.*.ppm [001-130] refers to the files football.001.ppm, football.002.ppm, ..., football.009.ppm, football.010.ppm, ..., football.130.ppm.

The third type of line is: single_star_expr [x-y+s], where the line is treated exactly as above, except that we skip by s. Thus, the line football.*.ppm [001-130+4] refers to the files football.001.ppm, football.005.ppm, football.009.ppm, football.013.ppm, etc.

Furthermore, a line may specify a shell command to execute to generate lines to be interpreted as described above, as if those lines were in the parameter file instead. Use back ticks, like in the Bourne Shell, like this:

    `cat myfilelist`

If input is from Standard Input (per the INPUT_DIR statement), ppmtompeg ignores the INPUT/END_INPUT block, but it still must be present.

ppmtompeg must convert all input files to one of the following formats as a first step of processing: PNM, YUV, JPEG(v4), or JMOVIE. (The conversion may be trivial if your input files are already in one of these formats). This line specifies which of the four formats. PPM is actually a subset of PNM. The separate specification is allowed for backward compatibility. Use PNM instead of PPM in new applications.
INPUT_CONVERT conversion_command
You must specify how to convert a file to the base file format. If no conversion is necessary, then you would just say:

Otherwise, conversion_command is a shell command that causes an image in the format your specified with BASE_FILE_FORMAT to be written to Standard Output. ppmtompeg executes the command once for each line between INPUT and END_INPUT (which is normally, but not necessarily, a file name). In the conversion command, ppmtompeg replaces each '*' with the contents of that line. If you had a bunch of gif files, you might say:

     INPUT_CONVERT giftopnm *
If you have a bunch of separate a.Y, a.U, and a.V files (where the U and V have already been subsampled), then you might say:
     INPUT_CONVERT cat *.Y *.U *.V

Input conversion is not allowed with input from stdin, so use

as described above.
SIZE widthxheight

width and height are the width and height of each frame in pixels.

When ppmtompeg can get this information from the input image files, it ignores the SIZE parameter and you may omit it.

When the image files are in YUV format, the files don't contain dimension information, so SIZE is required.

When ppmtompeg is running in parallel mode, not all of the processes in the network have access to the image files, so SIZE is required and must give the same dimensions as the input image files.

YUV_SIZE widthxheight
This is an obsolete synonym of SIZE.
This is meaningful only when BASE_FILE_FORMAT specifies YUV format, and then it is required. It specifies the sub-format of the YUV class.
n is the number of frames in a Group of Pictures. Except that because a GOP must start with an I frame, ppmtompeg makes a GOP as much longer than n as it has to to make the next GOP start with an I frame.

Normally, it makes sense to make your GOP size a multiple of your pattern length (the latter is determined by the PATTERN parameter file statement).

See Group Of Pictures.

n is roughly the number of slices per frame. Note, at least one MPEG player may complain if slices do not start at the left side of an image. To ensure this does not happen, make sure the number of rows is divisible by SLICES_PER_FRAME.
use half-pixel motion vectors, or just full-pixel ones It is usually important that you use half-pixel motion vectors, because it results in both better quality and better compression.
Use a search range of n pixels in each of the four directions from a subject pixel. (So the search window is a square n*2 pixels on a side).
This statement tells ppmtompeg what kind of search technique (algorithm) to use for P frames. You select the desired combination of speed and compression. EXHAUSTIVE gives the best compression, but LOGARITHMIC is the fastest. TWOLEVEL is an exhaustive full-pixel search, followed by a local half- pixel search around the best full-pixel vector (the PIXEL option is ignored for this search technique).
This statement tells ppmtompeg what kind of search technique (algorithm) to use for B frames. SIMPLE means find best forward and backward vectors, then interpolate. CROSS2 means find those two vectors, then see what backward vector best matches the best forward vector, and vice versa. EXHAUSTIVE does an n-squared search and is extremely slow in relation to the others (CROSS2 is about half as fast as SIMPLE).
Use n as the qscale for I frames. See Qscale.
Use n as the qscale for P frames. See Qscale.
Use n as the qscale for B frames. See Qscale.
This statement determines whether ppmtompeg uses the original images or the decoded images when computing motion vectors. Using decoded images is more accurate and should increase the playback quality of the output, but it makes the encoding take longer and seems to give worse compression. It also causes some complications with parallel encoding. (see the section on parallel encoding). One thing you can do as a trade-off is select ORIGINAL here, and lower the qscale (see QSCALE if the quality is not good enough.
Original or Decoded? (Normalized)
Reference Compression Speed Quality I Quality P Quality B
Decoded 1000 1000 1000 969 919
Original 885 1373 1000 912 884

The following lines are optional:

This statement is obsolete. It does nothing.

Before Netpbm 10.26 (January 2005), ppmtompeg would drop trailing B frames from your movie, since a movie can't end with a B frame. (See I Frames, P Frames, B Frames. You would have to specify FORCE_ENCODE_LAST_FRAME to stop that from happening and get the same function that ppmtompeg has today.

This statement specifies a custom non-intra quantization table. If you don't specify this statement, ppmtompeg uses a default non-intra quantization table.

The 8 lines immediately following NIQTABLE specify the quantization table. Each line defines a table row and consists of 8 integers, whitespace-delimited, which define the table columns.

This is analogous to NIQTABLE, but for the intra quantization table.
This statement specifies the aspect ratio for ppmtompeg to specify in the MPEG output. I'm not sure what this is used for.

ratio must be 1.0, 0.6735, 0.7031, 0.7615, 0.8055, 0.8437, 0.8935, 0.9157, 0.9815, 1.0255, 1.0695, 1.0950, 1.1575, or 1.2015.

This specifies the frame rate for ppmtompeg to specify in the MPEG output. Some players use this value to determine the playback rate.

rate must be 23.976, 24, 25, 29.97, 30, 50, 59.94, or 60.

This specifies the bit rate for Constant Bit Rate (CBR) encoding.

rate must be an integer.

This specifies the value ppmtompeg is to specify in the MPEG output for the Video Buffering Verifier (VBV) buffer size needed to decode the sequence.

A Video Verifying Buffer is a buffer in which a decoder keeps the decoded bits in order to match the uneven speed of the decoding with the required constant playback speed.

As ppmtompeg encodes the image, it simulates the decoding process in terms of how many bits would be in the VBV as each frame gets decoded, assuming a VBV of the size you indicate.

If you specify the WARN_VBV_UNDERFLOW statement, ppmtompeg issues a warning each time the simulation underflows the buffer, which suggests that an underflow would occur on playback, which suggests the buffer is too small.

If you specify the WARN_VBV_OVERFLOW statement, ppmtompeg issues a warning each time the simulation overflows the buffer, which suggests that an overflow would occur on playback, which suggests the buffer is too small.


These options were new in Netpbm 10.26 (January 2005). Before that, ppmtompeg issued the warnings always.

The following statements apply only to parallel operation:
This statement, paired with END PARALLEL, is what causes ppmtompeg to operate in parallel mode. See Parallel Operation.
This goes with PARALLEL.
The master starts off by measuring each slave's speed. It does this by giving each slave n frames to encode and noting how long the slave takes to finish. These are not just test frames, though -- they're real frames and the results become part of the output. ppmtompeg is old and measures time in undivided seconds, so to get useful timings, specify enough frames that it will take at least 5 seconds to process them. The default is 10.

If you specify FORCE_I_ALIGN, ppmtompeg will increase the test frames value enough to maintain the alignment.

If there aren't enough frames for every slave to have the indicated number of test frames, ppmtompeg will give some slaves fewer.

When you specify this statement, the master attempts to feed work to the slaves in chunks that take t seconds to process. It uses the speed measurement it made when it started up (see PARALLEL_TEST_FRAMES) to decide how many frames to put in the chunk. This statement obviously doesn't affect the first batch of work sent to each slave, which is the one used to measure the slave's speed.

Smaller values of t increase communication, but improve load balancing. The default is 30 seconds.


When you specify this statement, the master distributes work like with PARALLEL_TIME_CHUNKS, except that the master chooses the number of seconds for the chunks. It starts with a large number and, as it gets closer to finishing the job, reduces it. That way, it reduces scheduling overhead when precise scheduling isn't helpful, but still prevents a slave from finishing early after all the work has already been handed out to the other slaves, and then sitting idle while there's still work to do.


If this statement is present, ppmtompeg schedules on the assumption that each machine is about the same speed. The master will simply divide up the frames evenly between the slaves -- each slave gets the same number of frames. If some slaves are faster than others, they will finish first and remain idle while the slower slaves continue.

This has the advantage of minimal scheduling overhead. Where slaves have different speeds, though, it makes inefficient use of the fast ones. Where slaves are the same speed, it also has the disadvantage that they all finish at the same time and feed their output to the single Combine Server in a burst, which makes less efficient use of the Combine Server and thus can increase the total elapsed time.


RSH remote_shell_command
ppmtompeg executes the shell command remote_shell_command to start a process on another machine. The default command is rsh, and whatever command you specify must have compatible semantics. ssh is usually compatible. The command ppmtompeg uses is one like this: ssh remote.host.com -l username shellcommand.

Be sure to set up .rhosts files or SSH key authorizations where needed. Otherwise, you'll have to type in passwords.

On some HP machines, rsh is the restricted shell, and you want to specify remsh.

This statement forces each slave to encode a chunk of frames which is a multiple of the pattern length (see PATTERN). Since the first frame in any pattern is an I frame, this forces each chunk encoded by a slave to begin with an I frame.

This document used to say there was an argument to FORCE_I_ALIGN which was the number of frames ppmtompeg would use (and was required to be a multiple of the pattern length). But ppmtompeg has apparently always ignored that argument, and it does now.

This statement causes ppmtompeg not to delete the temporary files it uses to transmit encoded frames to the combine server. This means you will be left with a file for each frame, the same as you would get with the -frames option.

This is mostly useful for debugging.

This works only if you're using a shared filesystem to communicate between the servers.

This option was new in Netpbm 10.26 (January 2005).

Parameter File Notes

If you use the -combine_gops option, then you need to specify only the SIZE and OUTPUT values in the parameter file. In addition, the parameter file may specify input GOP files in the same manner as normal input files -- except instead of using INPUT_DIR, INPUT, and END_INPUT, use GOP_INPUT_DIR, GOP_INPUT, and GOP_END_INPUT. If you specify no input GOP files, then ppmtompeg uses by default the output file name with suffix .gop.gop_num, with gop_num starting from 0, as the input files.

If you use the -combine_frames option, then you need to specify only the SIZE, GOP_SIZE, and OUTPUT values in the parameter file. In addition, the parameter file may specify input frame files in the same manner as normal input files -- except instead of using INPUT_DIR, INPUT, and END_INPUT, use FRAME_INPUT_DIR, FRAME_INPUT, and FRAME_END_INPUT. If no input frame files are specified, then the default is to use the output file name with suffix .frame.frame_num, with frame_num starting from 0, as the input files.

Any number of spaces and tabs may come between each option and value. Lines beginning with # are ignored. Any other lines are ignored except for those between INPUT and END_INPUT. This allows you to use the same parameter file for normal usage and for -combine_gops and -combine_frames.

The file format is case-sensitive so all keywords should be in upper case.

The statements may appear in any order, except that the order within a block statement (such as INPUT ... END INPUT) is significant.

ppmtompeg is prepared to handle up to 16 B frames between reference frames when encoding with input from stdin. (To build a modified ppmtompeg with a higher limit, change the constant B_FRAME_RUN in frame.c and recompile).



The quantization scale values (qscale) give a trade-off between quality and compression. Using different Qscale values has very little effect on speed. The qscale values can be set separately for I, P, and B frames.

You select the qscale values with the IQSCALE, PQSCALE, and BSCALE parameter file statements.

A qscale value is an integer from 1 to 31. Larger numbers give better compression, but worse quality. In the following, the quality numbers are peak signal-to-noise ratio, defined as: signal-to-noise formula where MSE is the mean squared error.

Flower garden tests:
Qscale vs Quality
Qscale I Frames P Frames B Frames
1 43.2 46.3 46.5
6 32.6 34.6 34.3
11 28.6 29.5 30.0
16 26.3 26.8 28.6
21 24.7 25.0 27.9
26 23.5 23.9 27.5
31 22.6 23.0 27.3
Qscale vs Compression
Qscale I Frames P Frames B Frames
1 2 2 2
6 7 10 15
11 11 18 43
16 15 29 97
21 19 41 173
26 24 56 256
31 28 73 330

Search Techniques

There are several different motion vector search techniques available. There are different techniques available for P frame search and B frame search. Using different search techniques present little difference in quality, but a large difference in compression and speed.

There are 4 types of P frame search: Exhaustive, TwoLevel, SubSample, and Logarithmic.

There are 3 types of B frame search: Exhaustive, Cross2, and Simple. The recommended search techniques are TwoLevel and Logarithmic for P frame search, and Cross2 and Simple for B frame search. Here are some numbers comparing the different search methods:
P frame Motion Vector Search (Normalized)
Technique Compression1 Speed 2 Quality 3
Exhaustive 1000 1000 1000
SubSample 1008 2456 1000
TwoLevel 1009 3237 1000
Logarithmic 1085 8229 998
B frame Motion Vector Search (Normalized)
Technique Compression1 Speed2 Quality3
Exhaustive 1000 1000 1000
Cross2 975 1000 996
Simple 938 1765 991
 1Smaller numbers are better compression.  2Larger numbers mean faster execution.  3Larger numbers mean better quality.

For some reason, Simple seems to give better compression, but it depends on the image sequence.

Select the search techniques with the PSEARCH_ALG and BSEARCH_ALG parameter file statements.

Group Of Pictures (GOP)

A Group of Pictures (GOP) is a roughly independently decodable sequence of frames. An MPEG video stream is made of one or more GOP's. You may specify how many frames should be in each GOP with the GOP_SIZE parameter file statement. A GOP always starts with an I frame.

Instead of encoding an entire sequence, you can encode a single GOP. To do this, use the -gop command option. You can later join the resulting GOP files at any time by running ppmtompeg with the -combine_gops command option.


A slice is an independently decodable unit in a frame. It can be as small as one macroblock, or it can be as big as the entire frame. Barring transmission error, adding slices does not change quality or speed; the only effect is slightly worse compression. More slices are used for noisy transmission so that errors are more recoverable. Since usually errors are not such a problem, we usually just use one slice per frame.

Control the slice size with the SLICES_PER_FRAME parameter file statement.

Some MPEG playback systems require that each slice consist of whole rows of macroblocks. If you are encoding for this kind of player, if the height of the image is H pixels, then you should set the SLICES_PER_FRAME to some number which divides H/16. For example, if the image is 240 pixels (15 macroblocks) high, then you should use only 15, 5, 3, or 1 slices per frame.

Note: these MPEG playback systems are really wrong, since the MPEG standard says this doesn't have to be so.

Search Window

The search window is the window in which ppmtompeg searches for motion vectors. The window is a square. You can specify the size of the square, and whether to allow half-pixel motion vectors or not, with the RANGE and PIXEL parameter file statements.

I Frames, P Frames, B Frames

In MPEG-1, a movie is represented as a sequence of MPEG frames, each of which is an I Frame, a P Frame, or a B Frame. Each represents an actual frame of the movie (don't get confused by the dual use of the word "frame." A movie frame is a graphical image. An MPEG frame is a set of data that describes a movie frame).

An I frame ("intra" frame) describes a movie frame in isolation -- without respect to any other frame in the movie. A P frame ("predictive" frame) describes a movie frame by describing how it differs from the movie frame described by the latest preceding I or P frame. A B frame ("bidirectional" frame) describes a movie frame by describing how it differs from the movie frames described by the nearest I or P frame before and after it.

Note that the first frame of a movie must be described by an I frame (because there is no previous movie frame) and the last movie frame must be described by an I or P frame (because there is no subsequent movie frame).

Beyond that, you can choose which frames are represented by which types. You specify a pattern, such as IBPBP and ppmtompeg simply repeats it over and over throughout the movie. The pattern affects speed, quality, and stream size. Here is a chart which shows some of the trade-offs:
Comparison of I/P/B Frames (Normalized)
Frame Type Size Speed Quality
I frames 1000 1000 1000
P frames 409 609 969
B frames 72 260 919
(this is with constant qscale)

A standard sequence is IBBPBBPBBPBBPBB.

Select the sequence with the PATTERN parameter file statement.

Since the last MPEG frame cannot be a B frame (see above), if the pattern you specify indicates a B frame for the last movie frame of the movie, ppmtompeg makes it an I frame instead.

Before Netpbm 10.26 (January 2005), ppmtompeg instead drops the trailing B frames by default, and you need the FORCE_ENCODE_LAST_FRAME parameter file statement to make it do this.

The MPEG frames don't appear in the MPEG-1 stream in the same order that the corresponding movie frames appear in the movie -- the B frames come after the I and P frames on which they are based. For example, if the movie is 4 frames that you will represent with the pattern IBBP, the MPEG-1 stream will start with an I frame describing movie frame 0. The next frame in the MPEG-1 stream is a P frame describing movie frame 3. The last two frames in the MPEG-1 stream are B frames describing movie frames 1 and 2, respectively.

Specifying Input and Output Files

Specify the input frame images with the INPUT_DIR, INPUT, END_INPUT, BASE_FILE_FORMAT, SIZE, YUV_FORMAT and INPUT_CONVERT parameter file statements.

Specify the output file with the OUTPUT parameter file statement.


ppmtompeg can generate a variety of statistics about the encoding. See the -stat, -snr, -mv_histogram, -quiet, -no_frame_summary, and -bit_rate_info options.


You can run ppmtompeg on multiple machines at once, encoding the same MPEG stream. When you do, the machines are used as shown in the following diagram. We call this "parallel mode."


To do parallel processing, put the statement

in the parameter file, followed by a listing of the machines, one machine per line, then
Each of the machine lines must be in one of two forms. If the machine has filesystem access to the input files, then the line is:

machine user executable

The executable is normally ppmtompeg (you may need to give the complete path if you've built for different architectures). If the machine does not have filesystem access to the input files, the line is:

REMOTE machine user executable parameter file

The -max_machines command option limits the number of machines ppmtompeg will use. If you specify more machines in the parameter file than -max_machines allows, ppmtompeg uses only the machines listed first. This is handy if you want to experiment with different amounts of parallelism.

In general, you should use full path file names when describing executables and parameter files. This includes the parameter file argument on the original invocation of ppmtompeg.

All file names must be the same on all systems (so if e.g. you're using an NFS filesystem, you must make sure it is mounted at the same mountpoint on all systems).

Because not all of the processes involved in parallel operation have easy access to the input files, you must specify the SIZE parameter file statement when you do parallel operation.

The machine on which you originally invoke ppmtompeg is the master machine. It hosts a "combine server,", a "decode server," and a number of "i/o servers," all as separate processes. The other machines in the network (listed in the parameter file) are slave machines. Each hosts a single process that continuously requests work from the master and does it. The slave process does the computation to encode MPEG frames. It processes frames in batches identified by the master.

The master uses a remote shell command to start a process on a slave machine. By default, it uses an rsh shell command to do this. But use the RSH parameter file statement to control this. The shell command the master executes remotely is ppmtompeg, but with options to indicate that it is to perform slave functions.

The various machines talk to each other over TCP connections. Each machine finds and binds to a free TCP port number and tells its partners the port number. These port numbers are at least 2048.

Use the PARALLEL_TEST_FRAMES, PARALLEL_TIME_CHUNKS, and PARALLEL_PERFECT parameter file statements to control the way the master divides up work among the slaves.

Use the -nice command option to cause all slave processes to run "nicely," i.e. as low priority processes. That way, this substantial and long-running CPU load will have minimal impact on other, possibly interactive, users of the systems.  


Here is a look at ppmtompeg speed, in single-node (not parallel) operation:
Compression Speed
Machine Type Macroblocks per second1
HP 9000/755 280
DEC 3000/400 247
HP 9000/750 191
Sparc 10 104
DEC 5000 68
1A macroblock is a 16x16 pixel square

The measurements in the table are with inputs and outputs via a conventional locally attached filesystem. If you are using a network filesystem over a single 10 MB/s Ethernet, that constrains your speed more than your CPU speed. In that case, don't expect to get better than 4 or 5 frames per second no matter how fast your CPUs are.

Network speed is even more of a bottleneck when the slaves do not have filesystem access to the input files -- i.e. you declare them REMOTE.

Where I/O is the bottleneck, size of the input frames can make a big difference. So YUV input is better than PPM, and JPEG is better than both.

When you're first trying to get parallel mode working, be sure to use the -debug_machines option so you can see what's going on. Also, -debug_sockets can help you diagnose communication problems.



Table Of Contents