Monotonic Time - Externals

7 years ago by Niklas Keller — view source

unread

Morning Internals,

PHP doesn't currently have a way to access a monotonic time. There's a
feature request in the bug tracker: bugs [dot] php [dot] net/
bug.php?id=68029

A monotonic time is required to measure time intervals correctly, e.g. for
implementing event loops based on stream_select. microtime() is affected by
NTP and leap seconds and can therefore not be reliably used for measuring
time intervals.

Event loops based on extensions like libuv can take advantage of libuv's
own API for accessing a monotonic time.

Related: twitter [dot] com/danluu/status/815442765811023872
Related: blog [dot] cloudflare [dot] com/how-and-why-the-leap-secon
d-affected-cloudflare-dns/

Node.js does have a monotonic time and calls the function
process.hrtime: nodejs [dot] org/api/process.html#process_p
rocess_hrtime_time

There's an existing PECL extension named hrtime: php [dot] net/manual/
en/book.hrtime.php

I think we can reuse parts of it to implement a function in core to access
a monotonic time, I don't think we need to merge the extension into core.

What do you think about adding such a function to the PHP core?

Regards, Niklas

PS: No links, because stupid PHP.net spam filter ...

7 years ago by Niklas Keller — view source

unread

Resending previous mail, because mail distribution was broken...

Morning Internals,

PHP doesn't currently have a way to access a monotonic time. There's a
feature request in the bug tracker: bugs [dot] php [dot]
net/bug.php?id=68029

Event loops based on extensions like libuv can take advantage of libuv's
own API for accessing a monotonic time.

Related: twitter [dot] com/danluu/status/815442765811023872
Related: blog [dot] cloudflare [dot] com/how-and-why-the-leap-secon
d-affected-cloudflare-dns/

Node.js does have a monotonic time and calls the function
process.hrtime: nodejs [dot] org/api/process.html#process_p
rocess_hrtime_time

There's an existing PECL extension named hrtime: php [dot]
net/manual/en/book.hrtime.php

I think we can reuse parts of it to implement a function in core to access
a monotonic time, I don't think we need to merge the extension into core.

What do you think about adding such a function to the PHP core?

Regards, Niklas

PS: No links, because stupid PHP.net spam filter ...

7 years ago by Michael Wallner — view source

unread

Resending previous mail, because mail distribution was broken...

Morning Internals,

PHP doesn't currently have a way to access a monotonic time. There's a
feature request in the bug tracker: bugs [dot] php [dot]
net/bug.php?id=68029

A monotonic time is required to measure time intervals correctly, e.g. for
implementing event loops based on stream_select. microtime() is affected by
NTP and leap seconds and can therefore not be reliably used for measuring
time intervals.

...

There's an existing PECL extension named hrtime: php [dot]
net/manual/en/book.hrtime.php

I think we can reuse parts of it to implement a function in core to access
a monotonic time, I don't think we need to merge the extension into core.

What do you think about adding such a function to the PHP core?

+1?

Actually, it seems, that hrtime doesn't let you access the current
value, which is, well, quite a major use case, isn't it?

--
Regards,
Mike

7 years ago by Niklas Keller — view source

unread

Actually, it seems, that hrtime doesn't let you access the current
value, which is, well, quite a major use case, isn't it?

Yes, that seemed weird to me, too. Did we miss anything, Anatol?

Regards, Niklas

7 years ago by Anatol Belski — view source

unread

Hi,

-----Original Message-----
From: Niklas Keller [mailto:me@kelunik.com]
Sent: Thursday, January 5, 2017 2:27 AM
To: Michael Wallner mike@php.net; Anatol Belski anatol.php@belski.net
Cc: PHP Internals internals@lists.php.net
Subject: Re: [PHP-DEV] Fwd: Monotonic Time

Actually, it seems, that hrtime doesn't let you access the current
value, which is, well, quite a major use case, isn't it?

Yes, that seemed weird to me, too. Did we miss anything, Anatol?

Yeah, I thought I did that already, but well ... The ext was more about performance testing, at the start at least. I made some refactoring to it yesterday, so it is now possible to implement an arbitrary userspace timer using the HRTime\PerformanceCounter class. Maybe should add also HRTime\Timer, but it should first get a thought about some good API approach.

The portable parts are bundled files timer.[h|c] in the hrtime repo, supporting POSIX, Windows and Mac. They can be reused, or taken as an example, for sure.

Regards

Anatol

7 years ago by Niklas Keller — view source

unread

Actually, it seems, that hrtime doesn't let you access the current
value, which is, well, quite a major use case, isn't it?

Yes, that seemed weird to me, too. Did we miss anything, Anatol?

Yeah, I thought I did that already, but well ... The ext was more about
performance testing, at the start at least. I made some refactoring to it
yesterday, so it is now possible to implement an arbitrary userspace timer
using the HRTime\PerformanceCounter class. Maybe should add also
HRTime\Timer, but it should first get a thought about some good API
approach.

The portable parts are bundled files timer.[h|c] in the hrtime repo,
supporting POSIX, Windows and Mac. They can be reused, or taken as an
example, for sure.

Hi Anatol,

Do you think we should merge hrtime into core or add a simple function just
like microtime() to ext/standard?

Regards, Niklas

7 years ago by Leigh — view source

unread

Hi Anatol,

Do you think we should merge hrtime into core or add a simple function just
like microtime() to ext/standard?

Regards, Niklas

It would be great if it could be a simple function, however it is going to
be difficult to produce something cross-platform and meaningful with a
single function. (I know Anatol has written portable code, I haven't looked
at it but I assume it's wrapping the various structs used by i.e.
performance counters, hrtime, etc.)

I don't think we can get away from having an object-based implementation to
wrap those internal structs, but I do think this is a useful addition
regardless of how it's presented.

7 years ago by Niklas Keller — view source

unread

2017-01-13 15:06 GMT+01:00 Leigh leight@gmail.com:

Hi Anatol,

Do you think we should merge hrtime into core or add a simple function
just
like microtime() to ext/standard?

Regards, Niklas

It would be great if it could be a simple function, however it is going to
be difficult to produce something cross-platform and meaningful with a
single function. (I know Anatol has written portable code, I haven't looked
at it but I assume it's wrapping the various structs used by i.e.
performance counters, hrtime, etc.)

I don't think we can get away from having an object-based implementation to
wrap those internal structs, but I do think this is a useful addition
regardless of how it's presented.

With timer.c it's very easy to have it as one function, in fact it's a
simple static method currently doing exactly what I think it should
probably do in core as a function.

http://svn.php.net/viewvc/pecl/hrtime/trunk/hrtime.c?revision=341600&view=markup
< Line 416.

Any name suggestions?

Regards, Niklas

7 years ago by Anatol Belski — view source

unread

Hi,

-----Original Message-----
From: Niklas Keller [mailto:me@kelunik.com]
Sent: Friday, January 13, 2017 4:46 PM
To: Leigh leight@gmail.com
Cc: Anatol Belski anatol.php@belski.net; Michael Wallner mike@php.net;
PHP Internals internals@lists.php.net; Bob Weinand bwoebi@php.net;
Daniel Lowrey rdlowrey@php.net
Subject: Re: [PHP-DEV] Fwd: Monotonic Time

2017-01-13 15:06 GMT+01:00 Leigh leight@gmail.com:

Hi Anatol,

Do you think we should merge hrtime into core or add a simple
function
just
like microtime() to ext/standard?

Regards, Niklas

It would be great if it could be a simple function, however it is
going to be difficult to produce something cross-platform and
meaningful with a single function. (I know Anatol has written portable
code, I haven't looked at it but I assume it's wrapping the various structs used
by i.e.
performance counters, hrtime, etc.)

I don't think we can get away from having an object-based
implementation to wrap those internal structs, but I do think this is
a useful addition regardless of how it's presented.

With timer.c it's very easy to have it as one function, in fact it's a simple static
method currently doing exactly what I think it should probably do in core as a
function.

http://svn.php.net/viewvc/pecl/hrtime/trunk/hrtime.c?revision=341600&view=
markup
< Line 416.

The PerformanceCounter class from the hrtime ext allows to implement any time measurement directly in PHP. It provides the most low level units used by the system, the ticks. To convert ticks into actual time, the tick frequency is required. The frequency can be actually hidden from the implementation, or could be exposed to PHP. Exposing might be useful for several reasons, fe if one would want to compare different systems. Anyway, these are the base bricks.

Niklas, about merging into core - not sure it needs the StopWatch class, but maybe. The performance counter class, at the very low level, could have more chance. Taking timer.c as base should be fine, too, even there is a significant improved potential down to the ASM level later, though very platform specific. In first place what would make sense to know - what exact usage would it be? More specific use case info is what is needed to discuss and figure out the required implementation.

Regards

Anatol

7 years ago by Niklas Keller — view source

unread

Hi,

-----Original Message-----
From: Niklas Keller [mailto:me@kelunik.com]
Sent: Friday, January 13, 2017 4:46 PM
To: Leigh leight@gmail.com
Cc: Anatol Belski anatol.php@belski.net; Michael Wallner <mike@php.net
;
PHP Internals internals@lists.php.net; Bob Weinand bwoebi@php.net;
Daniel Lowrey rdlowrey@php.net
Subject: Re: [PHP-DEV] Fwd: Monotonic Time

2017-01-13 15:06 GMT+01:00 Leigh leight@gmail.com:

Hi Anatol,

Do you think we should merge hrtime into core or add a simple
function
just
like microtime() to ext/standard?

Regards, Niklas

It would be great if it could be a simple function, however it is
going to be difficult to produce something cross-platform and
meaningful with a single function. (I know Anatol has written portable
code, I haven't looked at it but I assume it's wrapping the various
structs used
by i.e.
performance counters, hrtime, etc.)

I don't think we can get away from having an object-based
implementation to wrap those internal structs, but I do think this is
a useful addition regardless of how it's presented.

With timer.c it's very easy to have it as one function, in fact it's a
simple static
method currently doing exactly what I think it should probably do in
core as a
function.

http://svn.php.net/viewvc/pecl/hrtime/trunk/hrtime.c?
revision=341600&view=
markup
< Line 416.

The PerformanceCounter class from the hrtime ext allows to implement any
time measurement directly in PHP. It provides the most low level units used
by the system, the ticks. To convert ticks into actual time, the tick
frequency is required. The frequency can be actually hidden from the
implementation, or could be exposed to PHP. Exposing might be useful for
several reasons, fe if one would want to compare different systems. Anyway,
these are the base bricks.

Niklas, about merging into core - not sure it needs the StopWatch class,
but maybe. The performance counter class, at the very low level, could have
more chance. Taking timer.c as base should be fine, too, even there is a
significant improved potential down to the ASM level later, though very
platform specific. In first place what would make sense to know - what
exact usage would it be? More specific use case info is what is needed to
discuss and figure out the required implementation.

Hi Anatol,

my specific use case is mainly for event loops, replacing the current
microtime calls there, e.g.
https://github.com/amphp/loop/blob/master/lib/NativeLoop.php#L62

Performance measuring is another one, but not that important to me. If
microtime drifts there, it will show wrong results, but it won't impact
applications with failures like executing DNS timeout watchers too early
and making things fail.

Regards, Niklas

7 years ago by Niklas Keller — view source

unread

I have implemented a hrtime() function which allows access to the
system's monotonic time in nanoseconds.

https://github.com/php/php-src/compare/master...kelunik:hrtime

Feedback is very welcome.

I'm not sure whether we should allow a parameter for the time unit or just
let the user do the simple calculation.

Regards, Niklas

7 years ago by Anatol Belski — view source

unread

Hi Niklas,

-----Original Message-----
From: Niklas Keller [mailto:me@kelunik.com]
Sent: Sunday, February 5, 2017 8:19 PM
To: Anatol Belski anatol.php@belski.net
Cc: Leigh leight@gmail.com; Michael Wallner mike@php.net; PHP Internals
internals@lists.php.net; Bob Weinand bwoebi@php.net; Daniel Lowrey
rdlowrey@php.net
Subject: Re: [PHP-DEV] Fwd: Monotonic Time

I have implemented a hrtime() function which allows access to the system's
monotonic time in nanoseconds.

Nice for the start.

https://github.com/php/php-src/compare/master...kelunik:hrtime

Feedback is very welcome.

I'm not sure whether we should allow a parameter for the time unit or just let
the user do the simple calculation.

I'd see this integration a bit different, as the goal is to put it into the core. Here some quick points to maybe think about

on the m4 side, symbols need to be checked
the PHP internal and user stuff should be integrated tight into the timer.c file,
a finer usage with the concrete platform should be taken care of, thus avoiding unnecessary function calls and conversions back/forth as you can use
there might be a situation, where the monotonic API could be used internally before modules are initialized
the symbols exported need to be renamed, probably prefix with php_monotonic_*, like php_monotonic_timer_current(), etc.
but aso, probably not all the symbols should be exported

With the names and API, probably some more clarity should be. AFM, the low level units should be exposed, rather than a concrete time. Also, the name hrtime() is probably not much speaking. Maybe these three?

sys_get_monotonic_ticks()
sys_get_monotonic_freq()
sys_get_monotonic_nanotime() or even - sys_get_monotonic_time(int $unit = NANOSECOND)

Regarding to "sys_get_*" scheme with these. But in general, probably nanoseconds were enough for the general case and fast, the pure ticks were for the advanced usage.

In general, probably for the core it should be done a more robust way, some parts on the original lib should get more error checks and improved. It would be probably more handy to discuss the further on the PR page.

Regards

Anatol

7 years ago by Fleshgrinder — view source

unread

With the names and API, probably some more clarity should be. AFM,
the low level units should be exposed, rather than a concrete time.
Also, the name hrtime() is probably not much speaking. Maybe these
three?

sys_get_monotonic_ticks()
sys_get_monotonic_freq()
sys_get_monotonic_nanotime() or even -
sys_get_monotonic_time(int $unit = NANOSECOND)

Regarding to "sys_get_*" scheme with these. But in general, probably
nanoseconds were enough for the general case and fast, the pure ticks
were for the advanced usage.

In general, probably for the core it should be done a more robust
way, some parts on the original lib should get more error checks and
improved. It would be probably more handy to discuss the further on
the PR page.

Regards

Anatol

Hey guys! :)

Please do not introduce more functions that require flags, it makes code
hard to read. It's much better to have dedicated functions that provide
exactly what is requested. A good bad example is microtime() vs
microtime(true). :(

sys_get_monotonic_nanotime() is rather long but at least descriptive.
Obviously providing a class could help with naming a lot and future
extension would be easy, e.g. MonotonicTime::getNanoseconds().

--
Richard "Fleshgrinder" Fussenegger

7 years ago by Stanislav Malyshev — view source

unread

Hi!

I think we can reuse parts of it to implement a function in core to access
a monotonic time, I don't think we need to merge the extension into core.

What's wrong with just having an extension?

--
Stas Malyshev
smalyshev@gmail.com

7 years ago by Niklas Keller — view source

unread

Hi!

I think we can reuse parts of it to implement a function in core to
access
a monotonic time, I don't think we need to merge the extension into core.

What's wrong with just having an extension?

Do you mean an external (not bundled) extension?

Simply not having it available by default. Measuring time frames is a
pretty fundamental operation, not only for event loops.

Regards, Niklas