19. Telemetry

In its cool and hot varieties, the MPS is capable of outputting a configurable stream of events (the telemetry stream) to assist with debugging and profiling.

The selection of events that appear in the stream is controlled by the environment variable MPS_TELEMETRY_CONTROL (by default none), and the stream is written to the file named by the environment variable MPS_TELEMETRY_FILENAME (by default mpsio.log).

The telemetry system writes blocks of binary output, and is fast enough to be left turned on in production code (the hot variety avoids emitting events on the critical path), which can be useful for diagnosing memory management problems in production environments.

The reporting of garbage collection statistics hasn’t always been suitable for deployment. John McCarthy described the first on-line demonstration of Lisp in an appendix to his paper “History of Lisp”:

Everything was going well, if slowly, when suddenly the Flexowriter began to type (at ten characters per second)

THE GARBAGE COLLECTOR HAS BEEN CALLED.
SOME INTERESTING STATISTICS ARE AS FOLLOWS:

and on and on and on. The garbage collector was quite new at the time, we were rather proud of it and curious about it, and our normal output was on a line printer, so it printed a full page every time it was called giving how many words were marked and how many were collected and the size of list space, etc. [...]

Nothing had ever been said about the garbage collector, and I could only imagine the reaction of the audience. We were already behind time on a tight schedule, it was clear that typing out the garbage collector message would take all the remaining time allocated to the demonstration, and both the lecturer and the audience were incapacitated with laughter. I think some of them thought we were victims of a practical joker.

19.1. Telemetry utilities

The telemetry system relies on three utility programs:

  • mpseventcnv decodes the machine-dependent binary event stream into a portable text format. It must be compiled for the same architecture as the MPS-linked program whose event stream it decodes.
  • mpseventtxt takes the output of mpseventcnv and outputs it in a more human-readable form.
  • mpseventsql takes the output of mpseventcnv and loads it into a SQLite database for further analysis.

You must build and install these programs as described in Building the Memory Pool System. These programs are described in more detail below.

19.2. Example

Here’s an example of turning on telemetry in the debugger and then encountering a corrupted object:

$ gdb ./scheme
GNU gdb 6.3.50-20050815 (Apple version gdb-1820) (Sat Jun 16 02:40:11 UTC 2012)
[...]
(gdb) set environment MPS_TELEMETRY_CONTROL=all
(gdb) run
Starting program: example/scheme/scheme
Reading symbols for shared libraries +............................. done
MPS Toy Scheme Example
[...]
7944, 0> (gc)
[...]
7968, 1> foo
Assertion failed: (TYPE(frame) == TYPE_PAIR), function lookup_in_frame, file scheme.c, line 1066.

Program received signal SIGABRT, Aborted.
0x00007fff91aeed46 in __kill ()

At this point there’s still output in the MPS’s internal event buffers, which needs to be flushed. It would be a good idea to add a call to mps_telemetry_flush() to the error handler, but for now we can just call it directly from the debugger:

(gdb) print mps_telemetry_flush()
$1 = void

The MPS writes the telemetry to the log in an encoded form for speed. It can be decoded using the mpseventcnv and mpseventtxt programs:

(gdb) shell mpseventcnv | sort | mpseventtxt > mpsio.txt

The sort is useful because the events are not necessarily written to the telemetry file in time order, but each event starts with a timestamp so sorting makes a time series. The decoded events look like this, with the timestamp in the first column, the event type in the second column, and then addresses or other data related to the event in the remaining columns. The source of the timestamp depends on the platform; it may be a low-cost high-resolution processor timer, such as the Time Stamp Counter on IA-32 and x86-64, if one is available. All numbers are given in hexadecimal.

000021C9DB3812C7 0075 EventClockSync      clock:0000000000001EE3
000021C9DB39E2FB 002B VMInit              vm:00007FFF5429C4B8 base:000000010BA4A000 limit:000000010BA4B000
000021C9DB3A5630 002D VMMap               vm:00007FFF5429C4B8 base:000000010BA4A000 limit:000000010BA4B000
000021C9DB3E6BAA 001A Intern              stringId:0000000000000002 string:"MFS"
000021C9DB3E6E17 001B Label               address:000000010BA0C5D8["MFS"] stringId:0000000000000002
000021C9DB3EB6F8 0044 PoolInitMFS         pool:000000010BA4A360 arena:000000010BA4A000 extendBy:0000000000001000 extendSelf:False unitSize:0000000000000030
000021C9DB3EFE3B 002B VMInit              vm:00007FFF5429C3D0 base:000000010BC84000 limit:000000010CC24000
000021C9DB3F33F3 002D VMMap               vm:00007FFF5429C3D0 base:000000010BC84000 limit:000000010BC85000

You can search through the telemetry for events related to particular addresses of interest.

In the example, we might look for events related to the address of the corrupted frame object:

(gdb) frame 3
#3  0x0000000100003f55 in lookup_in_frame (frame=0x1003fa7d0, symbol=0x1003faf20) at scheme.c:1066
1066            assert(TYPE(frame) == TYPE_PAIR);
(gdb) print frame
$2 = (obj_t) 0x1003fa7d0
(gdb) shell grep -i 1003fa7d0 mpsio.txt || echo not found
not found

There are no events related to this address, so in particular this address was never fixed (no TraceFix event).

Note

You may find it useful to add the command:

set environment MPS_TELEMETRY_CONTROL=all

to your .gdbinit.

19.3. Event categories

The “bit” column gives the bit number in the telemetry filter. These numbers are liable to change, but the current meanings (zero being the least significant bit) are:

Bit Name Description
0 Arena Per space or arena.
1 Pool Per pool.
2 Trace Per trace or scan.
3 Seg Per page (segment).
4 Ref Per reference or fix.
5 Object Per allocation, block, or object.
6 User User-invoked events: see mps_telemetry_intern().

19.4. Environment variables

In the ANSI plinth (the plinth that comes as default with the MPS), these two environment variables control the behaviour of the telemetry feature.

MPS_TELEMETRY_CONTROL

The event categories which should be included in the telemetry stream.

If its value can be interpreted as a number, then this number represents the set of event categories as a bitmap. For example, this turns on the Pool and Seg event categories:

MPS_TELEMETRY_CONTROL=6

Otherwise, the value is split into words at spaces, and any word that names an event category turns it on. For example:

MPS_TELEMETRY_CONTROL="arena pool trace"

The special event category all turns on all events.

MPS_TELEMETRY_FILENAME

The name of the file to which the telemetry stream should be written. Defaults to mpsio.log. For example:

MPS_TELEMETRY_FILENAME=$(mktemp -t mps)

In addition, the following environment variable controls the behaviour of the mpseventsql program.

MPS_TELEMETRY_DATABASE

The name of a SQLite database file that will be updated with the events from the decoded telemetry stream, if it is not specified with the -d option. If this variable is not assigned, mpsevent.db is used.

19.5. Decoding the telemetry stream

The MPS writes the telemetry stream in a binary encoded format for speed. The encoding is specific to the platform the program was running on, and so the output needs to be decoded before it can be processed.

The decoding takes place in two stages. First, the program mpseventcnv converts the binary encoded format into a portable text format suitable for input to one of the second-stage tools (mpseventtxt and mpseventsql).

-f <filename>

The name of the file containing the telemetry stream to decode. Defaults to mpsio.log.

-h

Help: print a usage message to standard output.

Note

mpseventcnv can only read telemetry streams that were written by an MPS compiled on the same platform.

Here’s some example output. The first column contains the timestamp of the event, the second column contains the event type, and remaining columns contain parameters related to the event.

000021C9DB3812C7   75 1EE3
000021C9DB39E2FB   2B 7FFF5429C4B8 10BA4A000 10BA4B000
000021C9DB3A5630   2D 7FFF5429C4B8 10BA4A000 10BA4B000
000021C9DB3E6BAA   1A 2 "MFS"
000021C9DB3E6E17   1B 10BA0C5D8 2
000021C9DB3EB6F8   44 10BA4A360 10BA4A000 1000 0 30
000021C9DB3EFE3B   2B 7FFF5429C3D0 10BC84000 10CC24000
000021C9DB3F33F3   2D 7FFF5429C3D0 10BC84000 10BC85000

19.6. Making the telemetry stream readable

The output of mpseventcnv can be made more readable by passing it through mpseventtxt, which takes the following options:

-l <filename>

The name of a file containing telemetry events that have been decoded by mpseventcnv. Defaults to standard input.

-h

Help: print a usage message to standard output.

For example, here’s the result of passing the output shown above through mpseventtxt:

000021C9DB3812C7 0075 EventClockSync      clock:0000000000001EE3
000021C9DB39E2FB 002B VMInit              vm:00007FFF5429C4B8 base:000000010BA4A000 limit:000000010BA4B000
000021C9DB3A5630 002D VMMap               vm:00007FFF5429C4B8 base:000000010BA4A000 limit:000000010BA4B000
000021C9DB3E6BAA 001A Intern              stringId:0000000000000002 string:"MFS"
000021C9DB3E6E17 001B Label               address:000000010BA0C5D8["MFS"] stringId:0000000000000002
000021C9DB3EB6F8 0044 PoolInitMFS         pool:000000010BA4A360 arena:000000010BA4A000 extendBy:0000000000001000 extendSelf:False unitSize:0000000000000030
000021C9DB3EFE3B 002B VMInit              vm:00007FFF5429C3D0 base:000000010BC84000 limit:000000010CC24000
000021C9DB3F33F3 002D VMMap               vm:00007FFF5429C3D0 base:000000010BC84000 limit:000000010BC85000

19.7. Loading the telemetry stream into SQLite

The decoded telemetry stream (as output by mpseventcnv) can be loaded into a SQLite database for further analysis by running mpseventsql.

mpseventsql takes the following options:

-i <filename>

The name of a file containing a decoded telemetry stream. Defaults to standard input.

-o <filename>

The name of a SQLite database file that will be updated with the events from the decoded telemetry stream specified by the -l option. The database will be created if it does not exist. If not specified, the file named by the environment variable MPS_TELEMETRY_DATABASE is used; if this variable is not assigned, mpsevent.db is used.

Updating a database with events from a file is idempotent unless the -f option is specified.

-d

Delete the database before importing.

-f

Forces the database to be updated with events from the decoded telemetry stream specified by the -i option, even if those events have previously been added.

-v

Increase the verbosity. With one or more -v options, mpseventsql prints informative messages to standard error. Verbosity levels up to 3 (-vvv) produce successively more detailed information.

This option implies -p.

-p

Show progress by printing a dot to standard output for every 100,000 events processed.

-t

Run internal tests.

-r

Rebuild the tables event_kind, event_type, and event_param. (This is necessary if you changed the event descriptions in eventdef.h.)

19.8. Telemetry interface

void mps_telemetry_flush(void)

Flush the internal event buffers into the telemetry stream.

This function also calls mps_io_flush() on the event stream itself. This ensures that even the latest events are now properly recorded, should the client program terminate (uncontrollably as a result of a bug, for example) or some interactive tool require access to the telemetry stream.

Note

Unless all arenas are properly destroyed (by calling mps_arena_destroy()), there are likely to be unflushed telemetry events when the program finishes. So in the case of abnormal program termination such as a fatal exception, you may want to call mps_telemetry_flush() explicitly.

mps_word_t mps_telemetry_get(void)

Return the telemetry filter.

void mps_telemetry_set(mps_word_t set_mask)

Set bits in the telemetry filter.

set_mask is a bitmask indicating the bits in the telemetry filter that should be set.

void mps_telemetry_reset(mps_word_t reset_mask)

Reset bits in the telemetry filter.

reset_mask is a bitmask indicating the bits in the telemetry filter that should be reset.

19.9. Telemetry labels

Telemetry labels allow the client program to associate strings with addresses in the telemetry stream. The string must first be interned by calling mps_telemetry_intern(), returning a label, and then the address can be associated with the label by calling mps_telemetry_label().

Typical uses of telemetry labels include:

  • labelling pools with a human-meaningful name;
  • labelling allocated objects with their type, class, or other description.

It is necessary to enable User events in the telemetry filter in order for telemetry labels to work. For example:

mps_label_t label;
mps_telemetry_set(1 << 6);
label = mps_telemetry_intern("symbol pool");
mps_telemetry_label(symbol_pool, label);

Labels are represented by the type mps_label_t. These are unsigned integers. After processing by mpseventsql, the association of addresses with labels appears in the EVENT_Label table, and the association of labels with strings appears in the EVENT_Intern table. These can then be used in queries, for example:

/* Pool name and creation time */
SELECT I.string, P.time
FROM EVENT_PoolInit AS P,
     EVENT_Label AS L,
     EVENT_Intern AS I
WHERE I.stringId = L.stringId AND L.address = P.pool;
mps_label_t mps_telemetry_intern(const char *label)

Registers a string with the MPS, and receives a telemetry label, suitable for passing to mps_telemetry_label().

label is a NUL-terminated string. Its length should not exceed 256 characters, including the terminating NUL.

Returns a telemetry label: a unique identifier that may be used to represent the string in future.

The intention of this function is to provide an identifier that can be used to concisely represent a string for the purposes of mps_telemetry_label().

Note

If the User event category is not turned on in the telemetry filter (via mps_telemetry_set() or MPS_TELEMETRY_CONTROL) then the string is not sent to the telemetry stream. A label is still returned in this case, but it is useless.

void mps_telemetry_label(mps_addr_t addr, mps_label_t label)

Associate a telemetry label returned from mps_telemetry_intern() with an address.

addr is an address.

label is a telemetry label returned from mps_telemetry_intern().

The label will be associated with the address when it appears in the telemetry stream.

Note

If the User event category is not turned on in the telemetry filter (via mps_telemetry_set() or MPS_TELEMETRY_CONTROL) then calling this function has no effect.

19.10. Customizing the telemetry system

If you need the telemetry system to support features not described here (for example, you need to transmit telemetry data over a network rather than writing it to a file on the local filesystem) then you may be able to do so by providing your own implementation of the I/O module.

When it first needs to output telemetry, the MPS call the plinth function mps_io_create() to create an I/O stream. It then calls mps_io_write() to write binary data to the stream and mps_io_flush() to flush the stream in response to mps_telemetry_flush(). By providing your own implementations of these functions, you can direct the telemetry stream wherever you like.

See Plinth for details.