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

There are four programs that help process telemetry streams:

  • 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.

  • mpseventpy emits Python data structures and constants for decoding a telemetry stream.

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.

000050C3BA05F734 0074 EventInit           major:2 median:3 minor:0 maxCode:143 maxNameLen:19 wordWidth:64 clocksPerSec:00000000000F4240
000050C3BA09FC24 0075 EventClockSync      clock:000000000000086C
000050C3BA0F22B7 002B VMInit              vm:00007FFEECCCB660 base:0000000103062000 limit:0000000103063000
000050C3BA0FA02F 008D GenInit             arena:0000000103062000 gen:00000001030624D8 serial:0 capacity:0000000000000400 mortality:   0.500
000050C3BA168B85 0044 PoolInitMFS         pool:0000000103062360 arena:0000000103062000 extendBy:0000000000001000 extendSelf:False unitSize:0000000000000030
000050C3BA168C3F 0015 PoolInit            pool:0000000103062360 arena:0000000103062000 poolClass:000000010301CD10 serial:0
000050C3BA16BB6F 002B VMInit              vm:00007FFEECCCB520 base:00000001032B3000 limit:000000010369B000
000050C3BA1787FC 0005 ArenaCreateVM       arena:0000000103062000 userSize:00000000003E8000 chunkSize:00000000003E8000 grainSize:0000000000001000 arenaClass:0000000103014DA8 serial:0

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.

000050C3BA05F734   74 2 3 0 8F 13 40 F4240
000050C3BA09FC24   75 86C
000050C3BA0F22B7   2B 7FFEECCCB660 103062000 103063000
000050C3BA0FA02F   8D 103062000 1030624D8 0 400 0.5
000050C3BA168B85   44 103062360 103062000 1000 0 30
000050C3BA168C3F   15 103062360 103062000 10301CD10 0
000050C3BA16BB6F   2B 7FFEECCCB520 1032B3000 10369B000
000050C3BA1787FC    5 103062000 3E8000 3E8000 1000 103014DA8 0

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:

000050C3BA05F734 0074 EventInit           major:2 median:3 minor:0 maxCode:143 maxNameLen:19 wordWidth:64 clocksPerSec:00000000000F4240
000050C3BA09FC24 0075 EventClockSync      clock:000000000000086C
000050C3BA0F22B7 002B VMInit              vm:00007FFEECCCB660 base:0000000103062000 limit:0000000103063000
000050C3BA0FA02F 008D GenInit             arena:0000000103062000 gen:00000001030624D8 serial:0 capacity:0000000000000400 mortality:   0.500
000050C3BA168B85 0044 PoolInitMFS         pool:0000000103062360 arena:0000000103062000 extendBy:0000000000001000 extendSelf:False unitSize:0000000000000030
000050C3BA168C3F 0015 PoolInit            pool:0000000103062360 arena:0000000103062000 poolClass:000000010301CD10 serial:0
000050C3BA16BB6F 002B VMInit              vm:00007FFEECCCB520 base:00000001032B3000 limit:000000010369B000
000050C3BA1787FC 0005 ArenaCreateVM       arena:0000000103062000 userSize:00000000003E8000 chunkSize:00000000003E8000 grainSize:0000000000001000 arenaClass:0000000103014DA8 serial:0

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. Decoding the telemetry stream in Python

mpseventpy takes no options, and emits Python code containing constants and data structures for decoding a telemetry stream generated by an application on the same platform and using the same version of the MPS.

To decode an event from a telemetry stream, start by reading and decoding the header.

HEADER_SIZE

Number of bytes in an event header. The event header consists of data that is common to all events, and precedes the event-specific data.

HEADER_FORMAT

Format string to pass to struct.unpack() to decode an event header.

class HeaderDesc

Named tuple describing an event header. It has the following attributes:

code is the code (an integer) for the event type.

size is the size of the remainder of event (in bytes).

clock is when the event occurred (in arbitrary time units).

Using these data structures, you might read an event from a file f like this:

header_data = f.read(HEADER_SIZE)
if not header_data:
    # No more telemetry.
header = HeaderDesc(*struct.unpack(HEADER_FORMAT, header_data))
event_data = f.read(header.size)
if not event_data:
    # Telemetry was truncated.

To decode the individual events, you’ll need the following data structures:

EVENT

Mapping from event code to EventDesc.

class EventDesc

Named tuple describing an event type. It has the following attributes:

name is the name of the event type.

code is the code (an integer) for the event type.

used is True if the event is used by the MPS, False if it is obsolete.

kind is the event category (see :py:Event categories), an instance of the KindDesc class.

params is a list of parameters of the event, each being an instance of the EventParam class.

maxsize is the maximum size of events of this type (in bytes).

format is a format string to pass to struct.unpack() to decode an event of this type.

class EventParam

Named tuple describing a parameter to an event type. It has the following attributes:

sort is a letter indicating the type of the parameter: P for a pointer to an internal MPS data structures, A for an address in the client program, W for a word, U for an unsigned integer, B for a Boolean, D for a double-precision floating-point number, and S for a string.

name is the name of the parameter.

doc is brief documentation for the parameter.

Using these data structures, you might decode an event like this:

event_desc = EVENT[header.code]
event_namedtuple = namedtuple(event_desc.name, [p.name for p in event_desc.params])
event = event_namedtuple(*struct.unpack(event_desc.format, event_data))

(In practice you’d want to cache the named tuple and reuse it for future events belonging to the same event type.)

19.9. Telemetry events

The set of telemetry events is not documented, and varies from version to version as we discover new requirements. You can see the current set of events by looking in the header eventdef.h.

If you have developed a tool that uses MPS telemetry, and would like to depend on particular telemetry events, contact us.

19.10. 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.

19.11. 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.

19.12. 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 the telemetry stream, the MPS calls 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.