[Pharo-project] Waste of CPU Power by polling events?
I am just porting squeak/Etoys to Samsung-PAD and i noticed some strange things: In Pharo 1.1 network connections seemed to be limited to 4 connections: In the listenLoop backlogSize is set to 4. Means - Pharo is limited to max. of four simultaneous TCP Connections? But: Some MacIntel VM seem to run Squeak with additional worker thread for event processing. Maybe this limitation may not occur here? Carbon ports of Pharo use ioPeekKeystroke, ioGetKeystroke, ioMousePoint, ioGetButtonState instead of kEventRawKeyModifiersChanged and kEventMouseMoved event Found in Pharo so far: isDraggingEvent isDropEvent isEditEvent isKbdEvent isMorphicEvent isNoteEvent isTempoEvent isWindowEvent Hmmm ... quite a lot of events being polled: Need additional: isOnMouseOverEvent isMouseGestureEvent isTouchGestureEvent isNetworkEvent isUSBEvent isSerialEvent isParallelPortEvents isCtrl-C-Event ... I fear, that this costs a lot of wasted VM Power by polling. UNIX, especially Linux is changing towards a event machine, driven by interrupts. So - Interpreters should also migrated from polling to interrupt driven event-machines. I am missing that in Pharo. Pharo still has too much polling code/algorithm running, that slows down Pharo/Squeak/Etoys ... unneccessarily. Any ideas appreciated ... tnx, regards Guido Stepken
Hi Guido, I have Samsung Galaxy Pad for a week now and I'm really interested to have a Smalltalk on it. So, you have all my moral support for your work! Janko On 22. 11. 2010 22:08, Guido Stepken wrote:
I am just porting squeak/Etoys to Samsung-PAD and i noticed some strange things:
-- Janko Mivšek AIDA/Web Smalltalk Web Application Server http://www.aidaweb.si
On Mon, 22 Nov 2010, Guido Stepken wrote:
I am just porting squeak/Etoys to Samsung-PAD and i noticed some strange things:
Are you porting Etoys to Pharo?
In Pharo 1.1 network connections seemed to be limited to 4 connections:
In the listenLoop backlogSize is set to 4. Means - Pharo is limited to max. of four simultaneous TCP Connections?
No. You're talking about ConnectionQueue which uses 4 as the size of the backlog queue. It means that at most 4 unaccepted connection attemts will allowed to wait, the rest will be rejected. This doesn't limit the number of accepted connections.
But: Some MacIntel VM seem to run Squeak with additional worker thread for event processing. Maybe this limitation may not occur here?
Carbon ports of Pharo use ioPeekKeystroke, ioGetKeystroke, ioMousePoint, ioGetButtonState instead of kEventRawKeyModifiersChanged and kEventMouseMoved event
Found in Pharo so far: isDraggingEvent isDropEvent isEditEvent isKbdEvent isMorphicEvent isNoteEvent isTempoEvent isWindowEvent
Hmmm ... quite a lot of events being polled:
Need additional:
isOnMouseOverEvent isMouseGestureEvent isTouchGestureEvent isNetworkEvent isUSBEvent isSerialEvent isParallelPortEvents isCtrl-C-Event ...
I fear, that this costs a lot of wasted VM Power by polling. UNIX, especially Linux is changing towards a event machine, driven by interrupts. So - Interpreters should also migrated from polling to interrupt driven event-machines.
I am missing that in Pharo. Pharo still has too much polling code/algorithm running, that slows down Pharo/Squeak/Etoys ... unneccessarily.
Any ideas appreciated ...
There are two classes for input event handling in Pharo 1.1: InputEventFetcher and InputEventPollingFetcher. You should be safe to use InputEventFetcher (which doesn't poll for events) on non-windows VMs, but the default is InputEventPollingFetcher. Levente
tnx, regards
Guido Stepken
On 22.11.2010 22:08, Guido Stepken wrote:
Need additional:
isOnMouseOverEvent isMouseGestureEvent isTouchGestureEvent You could try checking out how the iOS code does support for complex events. It's of course somewhat apple-specific, but hey, would be nice to have the same image-side representation and handling of touch-based events if possible.
I fear, that this costs a lot of wasted VM Power by polling. UNIX, especially Linux is changing towards a event machine, driven by interrupts. So - Interpreters should also migrated from polling to interrupt driven event-machines. I am missing that in Pharo. Pharo still has too much polling code/algorithm running, that slows down Pharo/Squeak/Etoys ... unneccessarily.
Any ideas appreciated ...
tnx, regards
Guido Stepken
At least for the input events, it is possible to avoid image-side polling if your VM signals the input semaphore. The Mac vm does it (haven't checked the unix vm, it might too), so if you're able to use that, you can swap out the InputEventPollingFetcher with InputEventFetcher to test how it works. Searching the list should give an example of how to do it properly, transfering existing handlers etc. With that, you should also be able to rewrite the polling Morphic loop to redraw as response to input handling that causes display invalidation, rather than process the available events as part of each draw step. Cheers, Henry
On 2010-11-23, at 1:23 AM, Henrik Sperre Johansen wrote:
On 22.11.2010 22:08, Guido Stepken wrote:
Need additional:
isOnMouseOverEvent isMouseGestureEvent isTouchGestureEvent You could try checking out how the iOS code does support for complex events. It's of course somewhat apple-specific, but hey, would be nice to have the same image-side representation and handling of touch-based events if possible.
Apple is very very notification &event callback oriented in their iOS platform. It sleeps until you touch, then you get a touch event, you process and go back to sleep. Heck if your startup time takes too many seconds Apple nails you. For Squeak on iOS I take the touch data place on a queue and signals the event semaphore. When the squeak image polls for the event we turn the touch data into a complex event and pass it back, then is processed by the Smalltalk code to turn into synthetic mouse events. (or something). For keyboard entry, we take the key input data and turn into synthetic keystroke events, and again signal the semaphore and put on a queue for later when the Smalltalk code wakes up and calls getnextevent to see what's up. I'm not sure if anyone mentioned the other issue is that the polling loop also deals with doing the morphic redrawing so it runs a cycle of doing I think it is 50 iterations a second. For WikiServer which doesn't show a squeak UI I could cut that back to doing a cycle every second or so. Oh and yet one more thing, the clock tick is yet another issue, some platforms assume they can run an interrupt 1000 a second (or more) to ensure the VM is responsive to Delay being correctly woken up to the microsecond. "PS here I am over simplify the issue & complexity of the issue" but on underpowered hardware the issues of how ioRelinquishProcessorForMicroseconds and the millisecond timer work become messy issues in relationship to CPU usage. -- =========================================================================== John M. McIntosh <johnmci@smalltalkconsulting.com> Twitter: squeaker68882 Corporate Smalltalk Consulting Ltd. http://www.smalltalkconsulting.com ===========================================================================
On Thu, Nov 25, 2010 at 8:25 PM, John M McIntosh < johnmci@smalltalkconsulting.com> wrote:
On 2010-11-23, at 1:23 AM, Henrik Sperre Johansen wrote:
On 22.11.2010 22:08, Guido Stepken wrote:
Need additional:
isOnMouseOverEvent isMouseGestureEvent isTouchGestureEvent You could try checking out how the iOS code does support for complex events. It's of course somewhat apple-specific, but hey, would be nice to have the same image-side representation and handling of touch-based events if possible.
Apple is very very notification &event callback oriented in their iOS platform. It sleeps until you touch, then you get a touch event, you process and go back to sleep. Heck if your startup time takes too many seconds Apple nails you.
For Squeak on iOS I take the touch data place on a queue and signals the event semaphore. When the squeak image polls for the event we turn the touch data into a complex event and pass it back, then is processed by the Smalltalk code to turn into synthetic mouse events. (or something).
For keyboard entry, we take the key input data and turn into synthetic keystroke events, and again signal the semaphore and put on a queue for later when the Smalltalk code wakes up and calls getnextevent to see what's up.
I'm not sure if anyone mentioned the other issue is that the polling loop also deals with doing the morphic redrawing so it runs a cycle of doing I think it is 50 iterations a second.
For WikiServer which doesn't show a squeak UI I could cut that back to doing a cycle every second or so.
Oh and yet one more thing, the clock tick is yet another issue, some platforms assume they can run an interrupt 1000 a second (or more) to ensure the VM is responsive to Delay being correctly woken up to the microsecond. "PS here I am over simplify the issue & complexity of the issue" but on underpowered hardware the issues of how ioRelinquishProcessorForMicroseconds and the millisecond timer work become messy issues in relationship to CPU usage.
Yes, we need to change this, but I think one has to change the background process and/or nothing to run behavior. In the VW VM the VM disables the heartbeat and enters some blocking call when there's nothing to do. Squeak runs the background process which calls a primitive that sleeps for a short time. The former is much better. If a delay is active as the VM does to sleep it schedules a wakeup event or supplies a timeout so that it blocks until input is available or the next delay. We could still keep the background process and do something similar to VW, disabling the heartbeat until after the relinquish, and factoring the delay into the relinquish timeout. Also the heartbeat already has facilities to change its frequency, so one could modify things to slow down the heartbeat, again ensuring that it ticks before the next delay expiry. What approach (not necessarily from the above) appeals to you John? best Eliot
-- =========================================================================== John M. McIntosh <johnmci@smalltalkconsulting.com> Twitter: squeaker68882 Corporate Smalltalk Consulting Ltd. http://www.smalltalkconsulting.com ===========================================================================
Eliot, I had looked at turn off the heartbeat when we did the nanosleep, then turned it back on. The problem was actually trying to figure out if it did any good since the cost of pausing/restarting the heartbeat task was high. I think first we have to quantify the costs. This is difficult to do on desktop machines since they are so fast, so I was working with an iPod touch which is 1000x (wild guess number) times slower. I'm also unsure of the pattern for when the ioRelinquishProcessorForMicroseconds is called and for how long. Further to that we do attempt to calculate the next wake up time, but sometime it's in the past, so why is ioRelinquishProcessorForMicroseconds being called? Even if we sleep then how long and what wakes us up early? On 2010-11-26, at 8:39 AM, Eliot Miranda wrote:
Yes, we need to change this, but I think one has to change the background process and/or nothing to run behavior. In the VW VM the VM disables the heartbeat and enters some blocking call when there's nothing to do. Squeak runs the background process which calls a primitive that sleeps for a short time. The former is much better. If a delay is active as the VM does to sleep it schedules a wakeup event or supplies a timeout so that it blocks until input is available or the next delay.
We could still keep the background process and do something similar to VW, disabling the heartbeat until after the relinquish, and factoring the delay into the relinquish timeout.
Also the heartbeat already has facilities to change its frequency, so one could modify things to slow down the heartbeat, again ensuring that it ticks before the next delay expiry.
What approach (not necessarily from the above) appeals to you John?
best Eliot
-- =========================================================================== John M. McIntosh <johnmci@smalltalkconsulting.com> Twitter: squeaker68882 Corporate Smalltalk Consulting Ltd. http://www.smalltalkconsulting.com ===========================================================================
In Hydra, i merged the heartbeat and delay timeout logic into single routine, which were running in high-priority thread. It simply picks the nearest possible time to wait (sleep) - either heartbeat cycle delay, or scheduled Delay timeout, then signals VM to interrupt and handle delay or do regular event processing using ioHandleEvents(). What is good in Windows, that there are special event mask for WaitForMultipleEvents() function, which could put a thread in wait till keyboard/mouse input available. In that way, on Windows, VM could avoid frequent interrupts to for polling events, and instead use waiting thread(s), and interrupt only if some event signaled. FreeBSD also introduced a kernel events, so instead of polling, one could simply wait for I/O activity (like file, socket etc). Not sure about mouse/keyboard input. I'm not sure, but think i heard linux kernel also integrated that recently. On 27 November 2010 00:40, John M McIntosh <johnmci@smalltalkconsulting.com> wrote:
Eliot, I had looked at turn off the heartbeat when we did the nanosleep, then turned it back on.  The problem was actually trying to figure out if it did any good since the cost of  pausing/restarting the heartbeat task was high. I think first we have to quantify the costs. This is difficult to do on desktop machines since they are so fast, so I was working with an iPod touch which is 1000x (wild guess number) times slower. I'm also unsure of the pattern for when the ioRelinquishProcessorForMicroseconds is called and for how long. Further to that we do attempt to calculate the next wake up time, but sometime it's in the past, so why is ioRelinquishProcessorForMicroseconds being called? Even if we sleep then how long and what wakes us up early? On 2010-11-26, at 8:39 AM, Eliot Miranda wrote:
Yes, we need to change this, but I think one has to change the background process and/or nothing to run behavior. Â In the VW VM the VM disables the heartbeat and enters some blocking call when there's nothing to do. Â Squeak runs the background process which calls a primitive that sleeps for a short time. Â The former is much better. Â If a delay is active as the VM does to sleep it schedules a wakeup event or supplies a timeout so that it blocks until input is available or the next delay. We could still keep the background process and do something similar to VW, disabling the heartbeat until after the relinquish, and factoring the delay into the relinquish timeout. Also the heartbeat already has facilities to change its frequency, so one could modify things to slow down the heartbeat, again ensuring that it ticks before the next delay expiry. What approach (not necessarily from the above) appeals to you John? best Eliot
-- =========================================================================== John M. McIntosh <johnmci@smalltalkconsulting.com>  Twitter:  squeaker68882 Corporate Smalltalk Consulting Ltd.  http://www.smalltalkconsulting.com ===========================================================================
-- Best regards, Igor Stasenko AKA sig.
On Fri, Nov 26, 2010 at 6:22 PM, Igor Stasenko <siguctua@gmail.com> wrote:
In Hydra, i merged the heartbeat and delay timeout logic into single routine, which were running in high-priority thread. It simply picks the nearest possible time to wait (sleep) - either heartbeat cycle delay, or scheduled Delay timeout, then signals VM to interrupt and handle delay or do regular event processing using ioHandleEvents().
What is good in Windows, that there are special event mask for WaitForMultipleEvents() function, which could put a thread in wait till keyboard/mouse input available. In that way, on Windows, VM could avoid frequent interrupts to for polling events, and instead use waiting thread(s), and interrupt only if some event signaled.
Yes, that's what I did in the VW VM. WaitForMultipleEvents also takes a timeout so one can wait until an event or the next delay with a single call.
FreeBSD also introduced a kernel events, so instead of polling, one could simply wait for I/O activity (like file, socket etc). Not sure about mouse/keyboard input. I'm not sure, but think i heard linux kernel also integrated that recently.
The problem with linux is that the pthreads implementation doesn't allow a normal user-level process to create high-priroity threads so that as soon as the Vm starts to spin doing some computation the heartbeat thread is shut-out and if the spinning computation only interrupted when a delay expires it'll never get interrupted because it is blocking the very thread that would signal the delay. So until linux's pthreads implementation supports multiple priorities we're stuck with hacks like the interval timer based itimer in the linux Cog VMs.
On 27 November 2010 00:40, John M McIntosh <johnmci@smalltalkconsulting.com> wrote:
Eliot, I had looked at turn off the heartbeat when we did the nanosleep,
then turned it back on. The problem was actually trying to figure out if
it did any good since the cost of pausing/restarting the heartbeat task was high. I think first we have to quantify the costs. This is difficult to do on desktop machines since they are so fast, so I was working with an iPod touch which is 1000x (wild guess number) times slower. I'm also unsure of the pattern for when the ioRelinquishProcessorForMicroseconds is called and for how long. Further to that we do attempt to calculate the next wake up time, but sometime it's in the past, so why is ioRelinquishProcessorForMicroseconds being called? Even if we sleep then how long and what wakes us up early? On 2010-11-26, at 8:39 AM, Eliot Miranda wrote:
Yes, we need to change this, but I think one has to change the background process and/or nothing to run behavior. In the VW VM the VM disables the heartbeat and enters some blocking call when there's nothing to do. Squeak runs the background process which calls a primitive that sleeps for a short time. The former is much better. If a delay is active as the VM does to sleep it schedules a wakeup event or supplies a timeout so that it blocks until input is available or the next delay. We could still keep the background process and do something similar to VW, disabling the heartbeat until after the relinquish, and factoring the delay into the relinquish timeout. Also the heartbeat already has facilities to change its frequency, so one could modify things to slow down the heartbeat, again ensuring that it ticks before the next delay expiry. What approach (not necessarily from the above) appeals to you John? best Eliot
--
===========================================================================
John M. McIntosh <johnmci@smalltalkconsulting.com> Twitter: squeaker68882 Corporate Smalltalk Consulting Ltd. http://www.smalltalkconsulting.com
===========================================================================
-- Best regards, Igor Stasenko AKA sig.
On Tue, 11 Jan 2011, Eliot Miranda wrote: (Pine can't quote your mail, sorry.) "The problem with linux is that the pthreads implementation doesn't allow a normal user-level process to create high-priroity threads so that as soon as the Vm starts to spin doing some computation the heartbeat thread is shut-out and if the spinning computation only interrupted when a delay expires it'll never get interrupted because it is blocking the very thread that would signal the delay. Â So until linux's pthreads implementation supports multiple priorities we're stuck with hacks like the interval timer based itimer in the linux Cog VMs." What about keeping the priority of the heartbeat thread at user-level and decreasing the other VM threads' priority slighly? Levente
2011/1/11 Levente Uzonyi <leves@elte.hu>
On Tue, 11 Jan 2011, Eliot Miranda wrote:
(Pine can't quote your mail, sorry.)
"The problem with linux is that the pthreads implementation doesn't allow a normal user-level process to create high-priroity threads so that as soon as the Vm starts to spin doing some computation the heartbeat thread is shut-out and if the spinning computation only interrupted when a delay expires it'll never get interrupted because it is blocking the very thread that would signal the delay. So until linux's pthreads implementation supports multiple priorities we're stuck with hacks like the interval timer based itimer in the linux Cog VMs."
What about keeping the priority of the heartbeat thread at user-level and decreasing the other VM threads' priority slighly?
Linux doesn't allow more than /one/ thread priority for threads in a user-level process. So one /can't/ have multiple priorities. All threads run at the same priority. This is a horrible bug in the linux pthreads implementation but there it is. best Eliot
Levente
On 11 January 2011 20:23, Eliot Miranda <eliot.miranda@gmail.com> wrote:
2011/1/11 Levente Uzonyi <leves@elte.hu>
On Tue, 11 Jan 2011, Eliot Miranda wrote:
(Pine can't quote your mail, sorry.)
"The problem with linux is that the pthreads implementation doesn't allow a normal user-level process to create high-priroity threads so that as soon as the Vm starts to spin doing some computation the heartbeat thread is shut-out and if the spinning computation only interrupted when a delay expires it'll never get interrupted because it is blocking the very thread that would signal the delay. Â So until linux's pthreads implementation supports multiple priorities we're stuck with hacks like the interval timer based itimer in the linux Cog VMs."
What about keeping the priority of the heartbeat thread at user-level and decreasing the other VM threads' priority slighly?
Linux doesn't allow more than /one/ thread priority for threads in a user-level process. Â So one /can't/ have multiple priorities. Â All threads run at the same priority. Â This is a horrible bug in the linux pthreads implementation but there it is.
So, why not use signals & POSIX timer [1] for same purpose? Or do signals having other issues which even worse than using heartbeat process? [1] http://www.kernel.org/doc/man-pages/online/pages/man2/timer_create.2.html
best Eliot
Levente
-- Best regards, Igor Stasenko AKA sig.
participants (7)
-
Eliot Miranda -
Guido Stepken -
Henrik Sperre Johansen -
Igor Stasenko -
Janko Mivšek -
John M McIntosh -
Levente Uzonyi