> I noticed that midas web pages consume unexpectedly large amount of resources, as observed by the chrome browser 
> "task manager" and by other tools.
> 
> For example, size of the "status" page was observe to reach 200, 600 and even 900 Mbytes.
> [this was fixed by using set_if_changed(e, v);
> 
> Also I observed the midas web pages consume an unusual amount of CPU - 5-10-15% - all in inactive tabs in minimized 
> windows.
> 

The case of high CPU use turned out to be quite nasty.

The symptoms:
- the "programs" page in an inactive tab in a minimized window sits "doing nothing" for a day or two.
- uses about 0 to 0.1 to 1% CPU and 40-50-60 Mbytes of RAM (after the previous improvements)
- suddenly I see it use 10-15-20% CPU, continuously, non stop
- I open this tab
- suddenly, CPU use goes to 100%, memory use quickly grows from 40-50-60 Mbytes to 100-200 Mbytes.
- after a few seconds everything settles down, CPU use is back to 0-0.1-1%, but memory use does not go down.
- WTH?!?

The culprit turned out to be the playing of the alarm sound. (I have all tabs "muted" by default, also speakers usually powered down).

If I comment-out the playing of the alarm sound, this problem goes away completely. Pretty conclusive, I think.

After adding lots of debug console.log() calls, I think I identified the problem: audio objects were being created,
but they were not starting to play their sound files. When I opened the tab, all of them (about 400) at the same time
loaded the mp3 file (resulting in memory use going from 50 Mbytes to 190 Mbytes, typical) and started playing
(as seen on the audio event activity in the cpu profile traces from the google-chrome "performance" tool).

I think I am looking at an unexpected interaction between audio objects and google-chrome throttling of inactive tabs.

To muddy the waters some more, google-chrome periodically fails audio.play() with an exception to the effect of
"we will not play audio because user is not interacting with this page enough". See
https://bitbucket.org/tmidas/midas/issues/191/exception-on-audioplay

Now I think I have this sort of fixed. I have to handle the audio.play() failure (which is not a normal exception,
but a rejected promise, the handler is quite different), and I do not allow creating new audio objects if previous
audio object did not finish playing.

(note the "normal" timing: periodic update every 1 sec, playing of alarm sound event 60 seconds, length of alarm sound file is 3 sec,
two sound files should never overlap. now a console.log message is printed if overlap is detected)

This leaves us with the problem of alarm sound not playing "because the user didn't interact with the document first",
and I think there is nothing I can do about that.

K.O.

P.S. Another quirk is I discovered: go to the "config" page and press the new buttons "play test sound" and "speak test message". In muted
tabs, the test sound will not sound, but the test message will be shouted out loudly. This seems inconsistent to me. Unwanted audio/video ads
are blocked but loud shouting of "shave with burma-shave" is permitted. I also wonder if speaking is subject to this
"user did not interact" business. If not, we could replace the playing of our relaxing alarm beep with the yelling of "alarm! alarm! alarm!".

K.O.

> > I noticed that midas web pages consume unexpectedly large amount of resources, as observed by the chrome browser 
> > "task manager" and by other tools.

The final fix is in. (plus a fix from Stefan).

When the audio.play() promise is rejected, one must clear audio.src, otherwise the browser will continue
with loading the audio file (but will not play it at the end).

Normally, this should not be a problem, but in inactive tabs, all activity is throttled down, and it so happens
that these audio objects accumulate (they are in the state of "we are trying to load the sound file, but
browser slows us down so much!"), consume huge amounts of memory (page memory use goes from ~50 Mbytes
to ~100-200 Mbytes) and consume huge amounts of CPU (not clear how, probably it's the firing of "loading", "canplay", etc
event handlers).

It does not help that mhttpd_fit_message() had a performance bug and consumed large amounts of CPU causing even
more slowing down by the be browser.

After adding audio.src="", all this is gone. I see no special CPU use, and I do not see any strange large memory use.

I still sometimes see inactive tabs grow from ~50 Mbytes to amount ~100 Mbytes. After I open them (activate them),
they quickly shrink back to ~50 Mbytes. I conclude that the browser is slowing down the garbage collector in inactive
tabs so much that it does not keep up with our 1/sec data polling.

So Stefan's fix to reduce polling from 1/sec to 1/10sec should help with this, too. (plus reduction of CPU use by fit_message() should
leave more time for the garbage collector to run).

P.S. General rules for browser slow down of inactive tabs seem to be written here:
https://developer.mozilla.org/en-US/docs/Web/API/Page_Visibility_API

Slowdown of timers is written here:
https://developer.mozilla.org/en-US/docs/Web/API/WindowOrWorkerGlobalScope/setTimeout#Notes

K.O.

> > > I noticed that midas web pages consume unexpectedly large amount of resources, as observed by the chrome browser 
> > > "task manager" and by other tools.

The work on this problem has been blogged in the bitbucket issue tracker:
https://bitbucket.org/tmidas/midas/issues/158/midas-status-page-memory-leak

K.O.

Below is a dump of the issue for posterity ---


Team Midas MIDAS related packages midas Issues
midas status page memory leak
Create issue
Issue #158 RESOLVEDOpen Workflow More Edit
dd1 created an issue 2018-12-25
I have the midas status page (https://daq16.triumf.ca/) open in macos google chrome 71.0.3578.98 and I watch in the "task manager" how the memory use is 
246 Mbytes and growing at around 1 Mbyte every 2-3 seconds. CPU use is around 3-5%, network use is 47 kBytes/sec. The slowly growing memory use 
indicates that we have a memory leak. (Note that javascript uses "automatic garbage collection" memory management, which does not eliminate memory 
leaks. Only capability to explicitly free unused memory is eliminated). K.O.

Comments (35)
dd1 REPORTER
Actual memory use goes up to around 250-something MBytes, then drops down to 240-something, them slowly grows back up, drops down, rinse, repeat. 
This is the javascript garbage collection in action. So there is no memory leak on the status page, but still why do we generate around 1 Mbyte/sec of 
javascript memory allocations? As comparison, the NYTimes front page consumes 270 Mbytes. One would expect the midas front page to be much more light 
weight... K.O.

Edit Pin to top Mark as spam Delete 2018-12-27
dd1 REPORTER
Then there is a question of memory use by the "message" page. This page does grow infinitely large by design - as new messages are added to midas.log - 
as as the user keeps scrolling the messages back in time. Perhaps we should somehow limit the total memory use there... K.O.

Edit Pin to top Mark as spam Delete 2018-12-27
Stefan Ritt
changed status to closed
I see the same behaviour. The relatively large memory allocation by Chrome probably comes from some bitmap caching. The browser prints the page 
contents into some temporary bitmap and then flushes it to the screen. That can easily take a few MB. I monitor such behaviour since several years now (for 
other processes) and concluded that I don't need to worry about JavaScript memory consumption.

Concerning the messages page: One line takes about 100 Bytes. If you scroll really fast, you can do maybe 30 lines per second, thus 3kB. If we allow the 
browser to consume another 100 MB (should be easily possible these days), you have to continuously scroll for 100000kb/3kb=30000 seconds or eight 
hours. Good luck!

Closing this topic if no complaints.

Pin to top Mark as spam Delete 2019-01-08
dd1 REPORTER
changed status to open
still see high memory use by midas pages. K.O.

Edit Pin to top Mark as spam Delete 2019-09-15
dd1 REPORTER
See high memory use from long running (days-weeks) web pages:

status page of my test experiment - 953 MB - 155 MB after reload
odb editor - 661 MB - 80 MB after reload
programs page - 602 MB - 64 MB after reload
sequencer - 253 MB - 151 MB after reload
sequencer - ??? (very big) - reloaded before I wrote it down
I think we are leaking memory somewhere. Or causing unnecessary allocations that the javascript garbage collector does not keep up with or does not 
cleanup correctly. K.O.

‌

‌

Edit Pin to top Mark as spam Delete 2019-09-15
dd1 REPORTER
I am suspicious of memory use trouble from periodic-update code that keeps setting innerHTML to the same value as it was before, unnecessarily. (this also 
causes other problems - cannot cut-and-paste affected parts of the web page, high cpu use to redraw the (unchanged) page). K.O.

‌

Edit Pin to top Mark as spam Delete 2019-09-15
Stefan Ritt
For setting innerHTML we should always use

if (text !== control.inner HTML)

control.innerHTML = text;

‌

I thought I caught most of the cases, but I might have missed some. Please add as needed.

‌

Stefan

Pin to top Mark as spam Delete 2019-09-15
dd1 REPORTER
Strange things continue. Just say huge CPU usage from 3 midas web pages (odb editor, programs page and the new sequencer page). All 3 pages are tabs in 
an iconized browser window. Suddenly machine feels slow, and I see all 3 use 25% CPU each (by the chrome-browser task manager window). Opened the 
browser window, sent to the offending tabs, nothing looks amiss, CPU usage went back to 0%. WTH? (all 3 pages have 100 Mbyte memory use, all 3 pages 
update at 1 Hz). K.O.

Edit Pin to top Mark as spam Delete 2019-09-16
dd1 REPORTER
looked at the “programs” page. learned how to use the google-chrome “performance” tool. I was definitely leaking html nodes. The leak was in an 
unexpected place - innerHTML with a link was miscomparing because of unexpected string transformation:

xbad: "<a href='?cmd=odb&odb_path=System/Clients/" + key + "'>" + host + "</a>";
good: "<a href=\"?cmd=odb&amp;odb_path=System/Clients/" + key + "\">" + host + "</a>";
Now node leak from my periodic update went from 35 nodes to 2 nodes per update. The performance tool fails to identify where these last 2 nodes are 
coming from.

K.O.

Edit Pin to top Mark as spam Delete 2019-09-17
dd1 REPORTER
Forgot to add - the periodic update from mhttpd_init() is also leaking nodes. I will look at it some other time. K.O.

Edit Pin to top Mark as spam Delete 2019-09-17
dd1 REPORTER
after improvement to the “programs” page, the tab is staying at 50-60 Mbytes. promising… K.O.

&#8204;

Edit Pin to top Mark as spam Delete 2019-09-18
dd1 REPORTER
Fixed node leak in mhttpd_refresh(): the alarm display was setting e.innerHTML even if it did not change.

There only remains an unavoidable node leak with “mheader_last_updated” where we set the current time every 1 second. If I comment this out, there is no 
node leak on the “programs” page.

K.O.

&#8204;

Edit Pin to top Mark as spam Delete 2019-09-18
dd1 REPORTER
“programs” page memory use now sits around 40 Mbytes. K.O.

Edit Pin to top Mark as spam Delete 2019-09-18
dd1 REPORTER
Stefan points me to the use of e.firstChild.data instead of e.innerHTML, per https://medium.com/@ok.bayat/fixing-memory-leak-problem-in-javascript-
application-ed3a2d9d92df

K.O.

Edit Pin to top Mark as spam Delete 2019-09-18
dd1 REPORTER
implemented this for the timestamp update and (i.e.) the “programs” page now leaks 0 nodes. memory use for all pages sits around 40-60 Mbytes. K.O.

Edit Pin to top Mark as spam Delete 2019-09-26
dd1 REPORTER
see problem of high cpu usage again, after google-chrome restarted after an update to latest version. for example, “program” page is 65 Mbytes, uses 20% 
CPU. (in an inactive tab). If I open this tab, for maybe 10 seconds, it goes to 100+ Mbytes with big CPU usage (>100%), then drops down to 90 Mbytes, 0% 
CPU usage. I do not see any other web pages or tabs doing this. Only our midas pages. WTH!?! K.O.

&#8204;

Edit Pin to top Mark as spam Delete 2019-10-13
dd1 REPORTER
figured out high cpu usage reported as “rendering”. Open “devtools”, goto “performance”, press Command-Shift-P, start typing “rendering”, select “fps 
meter”. A black square will open in top-left, showing graphics activity (frame rate, GPU usage, etc).

Now wait for new message to appear in the top status bar. It will be “yellow” at first, that it will fade to “gray”. During this fading, GPU use is 100% during 
about 1 second, FPS is about 50 frames/sec.

K.O.

Edit Pin to top Mark as spam Delete 2019-10-14
dd1 REPORTER
quick google search shows much discussion about css animations using “too much CPU”, i.e. google “css pulsing background”, but no clear way to tell the 
browser to slow down. It looks to me like the background-color animation tries to run at maximum possible frame-rate, as if electricity is free. (Since I am 
debugging high-cpu and high-memory use of inactive tabs, there is nobody looking at these animations). K.O.

Edit Pin to top Mark as spam Delete 2019-10-14
Stefan Ritt
New messages are displayed with a yellow background and fade to grey after 5 seconds. This is handeled in mhttpd.js around line 2144. You can try to 
remove the lines

d.style.setProperty("-webkit-transition", "background-color 3s", "");
d.style.setProperty("transition", "background-color 3s", "");
and see if the CPU load goes down.

Pin to top Mark as spam Delete 2019-10-14
dd1 REPORTER
captured another trace of midas page using 20% cpu in an inactive tab, iconized browser window. capturing is difficult, requires very fast mousing to: select 
the right tab, right-click to “inspect”, select “performance” tab, click on “start capture”, and hope that by this time the web page activity does not complete. 
this time I got the last 200 ms or so.

what I see is again is “media activity” (only identified as “task”), GPU activity (only identified as “GPU activity”) and main thread activity (identified as an 
infinitely repeating sequence of “receive response 206 audio/mpeg”, “receive data 39287 bytes”, “finish loading”, then the same sequence again. 39287 is 
the file size of resources/beep.mp3. There is no corresponding network activity, so the loading of beep.mp3 must be coming from cache. On the javascript 
console, there are the usual “not allowed to play audio because user did not interact” messages repeating about every 1-2-3 minutes.

I read this as: for reasons unknown, a huge number of audio requests becomes queued (the tab was inactive/iconized for many days) then they start trying to 
play (load beep.mp3, do not play it because “not allowed”, move on to the next audio object, load … etc). This is consistent with the cpu use, with the 
captured traces and with the quick growth in memory size (beep.mp3 objects are created, consume memory, cannot be free’d until garbage collector runs 
later. much later).

The above scenario is impossible with how the current audio playing code is written (only one audio object can exist at a time, new audio object can only be 
created after the previous one finished playing).

Two possible explanations: (a) the code running in the web page is not the same code as in mhttpd.js (running an old version from cache) or (b) the code 
“one audio object at time” is not working correctly if javascript code is throttled /delayed/stopped in inactive tabs.

Following code will have this problem:

var only_one = null;
function foo() { /* runs from periodic timer */
 if (!only_one) {
   a = new Audio(“beep.mp3”);
   /* throttled/suspended/delayed here */
   /* multiple Audio objects created because "only_one" is still null */
   only_one = a;
 }
}
K.O.

&#8204;

Edit Pin to top Mark as spam Delete 2019-10-17
dd1 REPORTER
fading background - yes, I found the code. pretty neat. I moved it around to remove the timer - I am suspicious of how the timers run in inactive tabs. but no 
time to study it.

but the current problem is clearly with audio objects, and the only audio we have is the periodic playing of beep.mp3. who knew there will be so much trouble.

there is still the unexplained use of GPU, but maybe playing/decoding mp3 files uses the GPU.

I am also puzzled why the status page from midas-2019-03 does not show any of these problems. it just sits there using no memory (50 Mbytes) and no 
CPU. perhaps we changed something in the playing of audio files since last March (when midas-2019-03 was tagged).

K.O.

&#8204;

Edit Pin to top Mark as spam Delete 2019-10-17
dd1 REPORTER
For the first time I saw my message “mhttpd_alarm_play: Cannot play alarm sound: previous alarm sound did not finish playing yet” reported on the javascript 
console. This confirms my guess that playing of audio is actually delayed and indeed we need to check that the previous audio finished playing before 
creating new audio objects. But the check in the current code has a race condition. If the delay/stall is inside “new Audio()”, we will create multiple audio 
objects as “last_audio” is still in the “finished playing” state, we only change it after the return from “new Audio()”. K.O.

Edit Pin to top Mark as spam Delete 2019-10-28
dd1 REPORTER
see big improvement. now inactive tabs grow from 50-ish Mbytes to 170-ish Mbytes, then when I open them, there is some cpu use (GC, I guess) and 
memory use drops back to 50-ish Mbytes. So we are not leaking any memory anymore. Looking at the console messages, I see that my fixes are helping - 
there is messages about attempts to create new Audio() when previous one did not finish yet. K.O.

Edit Pin to top Mark as spam Delete 2019-11-04
dd1 REPORTER
I guess, inactive tabs are throttled by google-chrome so much that their GC (memory garbage collection) does not keep up with our 1/sec data updates. I do 
not think we need to keep updating inactive tabs at this high frequency, but I am not sure how to detect if we are active or inactive. Maybe I can detect the 
throttling instead. K.O.

&#8204;

Edit Pin to top Mark as spam Delete 2019-11-04
dd1 REPORTER
see consistent behaviour from google-chrome:

have all these midas tabs open, inactive, window iconized, typical tab size is 50-ish Mbytes.
google-chrome update arrives
update is installed, all windows and tabs automatically closed, then reopened.
the midas tabs are still inactive, window is iconized
after a few days, see behaviour as described before:
midas tabs use 20-30% CPU, size is 100-ish Mbytes
if I open one of these tabs, it’s cpu usage goes up to 160%, size grows to 250-ish Mbytes, then within 5-10 seconds drops to 100 Mbytes, CPU usage goes 
from 160% to zero.
when looking at this, if I am quick enough, I can right-click “inspect”, go to the “performance” tab, and press the “start collecting data” button and I capture 
the very tail end of all this strange activity. This is the traces I have been describing so far.
K.O.

Edit Pin to top Mark as spam Delete 2019-11-07
dd1 REPORTER
see big blob of activity:

timer activation
mhttpd_message()
first call to mhttpd_fit_message()
a long cycle of (maybe 10-20) “recalculate size”, “layout”, “parse html”
second call to mhttpd_fit_message()
same long cycle of …
The way I understand this, mhttpd_fit_message() changes the size of some html element that causes the whole window to be re-layed-out.

K.O.

&#8204;

Edit Pin to top Mark as spam Delete 2019-11-07
dd1 REPORTER
trying to figure out what triggers a long run of the “rasterizer” thread. see a very strange call sequence

timer fires
mhttpd_alarm_play()
mhttpdConfig()
“Layout”
mhttpd_alarm_play() calls mhttpdConfig() (3 times) to find out if alarm sound is enabled, the period, the file name, etc. so far so good. but mhttpdConfig() 
does not touch any DOM objects, so why is it shown as calling “Layout”?!?

other than this trace, I see nothing else that would trigger the rasterizer thread…

(note) this time, mhttpd_alarm_play() does not call mhttpd_alarm_play_now(), so “new Audio” and stuff does not enter this picture.

K.O.

Edit Pin to top Mark as spam Delete 2019-11-07
dd1 REPORTER
in the early part of the trace, where I think the meat of “tab is using cpu and memory” is, I see the audio events firing in rapid sequence: loadeddata, canplay, 
canplaythrough, rinse, repeat.

It turns out that promise rejection from audio.play() does not stop the loading of the sound file. This is easy to see by attaching the event handlers to these 
events and by observing these event handlers print something to the javascript console.

If that is what is happening, it explains what I see: all my previous attempts to prevent the piling up of sound files are unsuccessful, and when I open the 
previously inactive tab, all the queued sound files start loading (and not playing per “user did not interact” policy).

google docs suggest using audio.src=”” to cancel loading of sound files and it does seem to work. testing it now.

K.O.

&#8204;

Edit Pin to top Mark as spam Delete 2019-11-07
dd1 REPORTER
gotcha. came back home, found one tab using about 10% cpu. audio.src=”” is commented out, javascript console is full of this:

Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_loadeddata: counter 234
mhttpd.js:2763 Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_canplay: counter 234
mhttpd.js:2767 Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_canplaythrough: counter 234
mhttpd.js:2759 Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_loadeddata: counter 235
mhttpd.js:2763 Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_canplay: counter 235
mhttpd.js:2767 Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_canplaythrough: counter 235
mhttpd.js:2759 Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_loadeddata: counter 236
mhttpd.js:2763 Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_canplay: counter 236
mhttpd.js:2767 Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_canplaythrough: counter 236
mhttpd.js:2759 Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_loadeddata: counter 237
mhttpd.js:2763 Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_canplay: counter 237
mhttpd.js:2767 Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_canplaythrough: counter 237
mhttpd.js:2759 Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_loadeddata: counter 238
mhttpd.js:2763 Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_canplay: counter 238
mhttpd.js:2767 Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_canplaythrough: counter 238
mhttpd.js:2759 Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_loadeddata: counter 239
mhttpd.js:2763 Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_canplay: counter 239
mhttpd.js:2767 Thu Nov 07 2019 22:17:30 GMT-0800 (Pacific Standard Time): mhttpd_audio_canplaythrough: counter 239
the timestamp is exactly when I opened this tab. so confirmed, a whole bunch of audio files got queued, when I open the tab they all try to play. (there is no 
actual sound, all tabs are muted).

now I uncomment audio.src=”” and see what happens.

K.O.

&#8204;

Edit Pin to top Mark as spam Delete 2019-11-07
dd1 REPORTER
looks good. an update to google-chrome came in and after installing, I no longer see midas tabs show high cpu usage or high memory use. I think the 
audio.src=”” fix is it. I will be committing these fixes to midas. K.O.

Edit Pin to top Mark as spam Delete 2019-11-12
Stefan Ritt
The loop in mhttpd_fit_message() is there for a good reason: I want to display the message in a single line. If it’s too long, I want first to cut the time stamp 
and then display it. If it’s still too long, I want to truncate the message and display “…” at the end. The problem is what “too long” is. Nobody can tell you how 
much pixel a message on your browser take, because this depends on the installed fonts, the exact character spacing of your browser and so on. So the only 
way I could make this happen is to add one char at a time, until we get close to the maximum allowed space. If course this requires a re-layout of the page for 
10-20 times, but when your window is in the foreground this is not a problem, since a browser can do this with small CPU load. The “scope” application I use 
does 70 frames per second at 30% CPU load. So one could make the loop a bit smarter, like binary search, which would drop the 10-20 iterations to log2(10-
20) ~ 4-5, but still there would be a loop.

How that the update of the messages in the background is suppressed with the hidden API, do you still have that problem or can we consider it fixed?

Stefan

Pin to top Mark as spam Delete 2019-11-18
dd1 REPORTER
see new behaviour - after many days, inactive page size is ~180 Mbytes. 0% CPU use (an improvement from before where there was large CPU use). activate 
the tab, nothing much happens, 0% CPU use (again an improvement from before). after about 30 seconds, memory use drops down to the normal 50-70-80 
Mbytes. I think what we see is the garbage collector is throttled down and does not keep up with our allocations. Stefan’s new fix reducing polling in inactive 
pages from 1/sec to 1/10sec should help with this. K.O.

&#8204;

Edit Pin to top Mark as spam Delete 2019-11-19
dd1 REPORTER
mhttpd_fit_message() - confirmed. I was confused about the function argument.

I thought it is passed an array of messages. no, it is one message string and the loop is over the message string length. The loop is done twice (second time, 
with the time/date stamp removed). google-chrome debugger does show that this uses large amount of CPU, mainly to compute d.offsetWidth.

I think I will refactor these loops - instead of growing the message, I will shrink it.

K.O.

Edit Pin to top Mark as spam Delete 4 hours ago
dd1 REPORTER
rewrote mhttpd_fit_message() to reduce CPU use: try to fit complete message, if too long, try to fit message without timestamp, if too long, guess the desired 
length assuming all chars have same width, then grow or shrink the message until the size is right. K.O.

Edit Pin to top Mark as spam Delete 17 minutes ago
dd1 REPORTER
changed status to resolved
The main fix is to set audio.src="" in the promise rejection. K.O.

Edit Pin to top Mark as spam Delete just now

What would you like to say?
Assignee
–
Type
bug
Priority
major
Status
resolved
Votes
0 Vote for this issue
Watchers
1 Stop watching
Dismiss this bannerJira Software: the preferred issue tracker for Bitbucket. Join the team!

I noticed that midas web pages consume unexpectedly large amount of resources, as observed by the chrome browser 
"task manager" and by other tools.

For example, size of the "status" page was observe to reach 200, 600 and even 900 Mbytes. The "programs" page (which 
does not have nearly as much stuff as the status page), was observed to reach 200-600 Mbytes. This is comparable to the 
New York Times front page, which has much more stuff, but usually runs at about 200 Mbytes. (they do force a periodic full 
page reload, to deal with exactly this same type of trouble, I suspect).

Also I observed the midas web pages consume an unusual amount of CPU - 5-10-15% - all in inactive tabs in minimized 
windows.

All this was quite noticeable in my oldish mac laptop with only 8 GBytes of RAM.

Using the google-chrome performance analyzer I was able to identify the reason of high memory use - our 1/sec periodic 
updates leak "too many" DOM "nodes" and I suspect that due to throttling of inactive tabs, the garbage collector simply 
does not keep up with us.

(Note that javascript features automatic memory management with garbage collection. In practice in means that where in 
C/C++ we have malloc() and free(), in javascript we only have malloc() and no free(), and cannot explicitly release memory 
we know we no longer need. In the C/C++ sense, all memory allocations are leaked, and one relies on a janitor to "clean it all 
up" eventually, later).

The source of node leakage was unexpected (unexpected to me). It turns out that each assignment to e.innerHTML creates 
a new node, even if the new contents is the same as the old contents. (also the html parser has to run, consuming extra cpu 
cycles).

Obvious solution is to write code like this:
if (v !== e.innerHTML) { e.innerHTML = v };

This helped quite a bit on the "programs" page, but not as much as expected, and hardly at all on the "status" page.

It turns out, that read of innerHTML does not necessarily return the same string as it was written into it.
For example, if "v" is "a&b", e.innerHTML will return "a&b" and the comparison will misfire.
There is more cases like this, see the section "Test set and get e.innerHTML" on the "example" midas page.

To help dealing with this, I suggest that instead of "inline" comparison (as above), one writes this:
mhttpd_set_if_changed(e, v);

Then to check that the comparison is effective, go to mhttpd.js and uncomment the console.log() call in 
mhttpd_set_if_changed(), reload the page and look at the javascript console to see all calls that result
in assignment of innerHTML (and leakage of DOM nodes).

This done, after replacing many "&" with "&" and many "\'" with "\"", node leakage on the "programs" page was reduced 
to 1 node per 1/sec update: the unavoidable change to the timestamp on the top-right of the page.

Luckily, Stefan pointed me to the solution for this: use of e.firstChild.data instead of e.innerHTML. The only quirk is that the 
node should not be empty, which was easy to arrange by setting the initial value of the timestamp to a dummy value.

With these changes, the "programs" page (and most other pages) now leak 0 nodes (from the 1/sec periodic updates). 

There is still some small memory leakage from making the RPC requests and from receiving the RPC replies, but the 
garbage collector seems to have no trouble with them.

Typical memory use for all midas pages is now 50-60 Mbytes (down from 100-200 Mbytes).

The "status" page took a bit more work to fix due to it's curious coding, but it, too now uses 50-60 Mbytes as well. It still 
leaks quite a few nodes (to be fixed!), but the garbage collector seems to keep up with the allocations.

K.O.

I reworked the alarm sounds in mhttpd - now you can turn off all sounds without disabling the 
alarm system for everybody.

a) new checkbox on the "alarms" page to turn off the alarm buzzer sound
b) fixed a bug where the status page will speak the last alarm even if the "speak" checkbox is 
unchecked on the "alarms" page (was coming through the TALK messages)
c) made sure the chat messages are only spoken if "speak" is enabled on the "chat" page
d) these speech and sounds settings are now stored in the browser "localStorage", which means 
they are shared across all open tabs and windows and are preserved across browser sessions and 
computer reboots.

I hope this is an improvement.

There is still one bug remaining - the first (last?) alarm is always spoken twice - 1st time in the loop 
over all alarms and 2nd time through the TALK messages. I do not know how to fix this.

K.O.

The mhttpd json-rpc interface now implements batch requests per
http://www.jsonrpc.org/specification#batch

In the nutshell, instead of a single request, one can send a json array of requests and receive a json 
array of replies.

As a variance from the spec, the midas implementation executes the requests strictly in-order and 
the array of replies corresponds exactly to the array of requests (the spec requires user to use the 
"id" field to match replies to requests, in midas json-rpc, the 1st reply is always to the 1st request,
2nd reply is to the 2nd request and so forth).

See this in action look at resources/example.html and in resources/transition.html

K.O.

I'm not sure if these issues are related or not, but I'm getting an error
message when I want to access the root of the ODB via the webserver:
[mhttpd,ERROR] [mhttpd.cxx:563:rread,ERROR] Cannot read file '/root', read of
4096 returned -1, errno 21 (Is a directory)

I also tried turning the re-direct from http to https off, but this does not
seem to work. I also noticed that the redirect changes the localhost into a
hostname. Where does mongoose take this hostname from?

EDIT: Seems that the change of the hostname is due to a setting in /etc/hosts,
i.e. all my fault ...

EDIT: I think there was some issue with the mhttpd. When I checked the output (I
used screen to run it), it was full of these messages:

ss_semaphore_wait_for: semop/semtimedop(2588679) returned -1, errno 43
(Identifier removed)
al_check: Something is wrong with our semaphore, ss_semaphore_wait_for()
returned 408, aborting.
al_check: Cannot abort - this will lock you out of odb. From this point, MIDAS
will not work correctly. Please read the discussion at
https://midas.triumf.ca/elog/Midas/945

Restarted it and it stopped redirecting. So accessing the root of the ODB via
the webserver is the only issue now.

> I'm not sure if these issues are related or not, but I'm getting an error
> message when I want to access the root of the ODB via the webserver:
> [mhttpd,ERROR] [mhttpd.cxx:563:rread,ERROR] Cannot read file '/root', read of
> 4096 returned -1, errno 21 (Is a directory)

This is an old bug. It was part of the "custom path" confusion. Fixed (I think) in all midas-2019 
releases.

To confirm, which version are you using (run "odbedit ver" or look on the mhttpd "help" page)?

If you have an older version, I recommend that you update to midas-2019-03 (cd midas; git pull; 
git checkout midas-2019-03; make clean; make).

If you feel adventurous, you can also update to the head of the development version
and see all the new features (cmake, c++11, new history pages).

If you do not feel adventurous, wait until we have midas-2019-09 ready, use midas-2019-03 
until then.

K.O.

Please, tell me how to choose a value of a "key" like DCM, pulser period,
presamples, upper thresholds to run a experiment? where can I find the related
informations? 

I am having problems with handling the end-of-run situation in my midas
frontend. I have a device that continuously sends data (over USB) and I read
this data in my "read_event" function.

Everything is good until the end-of-run, at which time this happens:
0) mfe.c calls my read_event() to read the data (loop until the end-of-run
transition)
1) mfe.c calls my end_of_run()
2) here, I tell the device "please stop sending data"
3) all seems good, but wait!!!
4) there is all this data generated between step 0 and step 2 still sitting
inside the device and it has nowhere to go: the run is ended, the output file is
closed, my read_event() will never be called ever again (well, until the next run).

It seems to me mfe.c needs to have one more function, something like
"pre_end_of_run()" that works like this:
0) mfe.c calls my read_event() to read the data (loop until the end-of-run
transition)
1) mfe.c calls pre_end_of_run(), here I tell the device to stop sending data
2) mfe.c calls read_event() for the very last time, to give me the opportunity
to read and send away any data I still may have.
3) mfe.c calls the end_of_run(). The run is truely finished.

Any thoughts?

K.O.

> I am having problems with handling the end-of-run situation in my midas
> frontend. I have a device that continuously sends data (over USB) and I read
> this data in my "read_event" function.
> 
> Everything is good until the end-of-run, at which time this happens:
> 0) mfe.c calls my read_event() to read the data (loop until the end-of-run
> transition)
> 1) mfe.c calls my end_of_run()
> 2) here, I tell the device "please stop sending data"
> 3) all seems good, but wait!!!
> 4) there is all this data generated between step 0 and step 2 still sitting
> inside the device and it has nowhere to go: the run is ended, the output file is
> closed, my read_event() will never be called ever again (well, until the next run).
> 
> It seems to me mfe.c needs to have one more function, something like
> "pre_end_of_run()" that works like this:
> 0) mfe.c calls my read_event() to read the data (loop until the end-of-run
> transition)
> 1) mfe.c calls pre_end_of_run(), here I tell the device to stop sending data
> 2) mfe.c calls read_event() for the very last time, to give me the opportunity
> to read and send away any data I still may have.
> 3) mfe.c calls the end_of_run(). The run is truely finished.
> 
> Any thoughts?

You can achieve the desired functionality without changing mfe.c:

0) mfe.c calls read_event
1) mfe.c calls end_of_run. Your end_of_run tells the device to stop data and flushes
the remaining data. At this point you have to re-make actually a part of the mfe.c
functionality, but basically you need a bm_compose_event() and a bm_send_event(), so
just a few lines of code. If you want to have the final event number right in your
equipment, you also need to update eq->events_sent accordingly. 

Given the fact that 99% of the experiments do not need this functionality, I propose
that we keep mfe.c and you add the few lines of code into your user part of the
specific frontend.

Stefan

[Dear All, I apologize in advance for spamming.]

1) I tried to login into the forum from the lab computer and realized 
   that I forgot my password
2) I tried to reset the password and found that when registering 
   I mistyped my email address, having typed '.giv' instead of '.gov'
   in the domain name, so the recovery email went into nowhere 

(still have one session open on the laptop so can post this question)

- how do I get my email address fixed so I'd be able to reset the password?

-- many thanks, Pasha

The code for storing midas history in an odbc sql database has been committed.
Changes:
include/history_odbc.h, src/history_odbc.cxx --- implementation
src/mlogger.c --- call the history_odbc functions
utils/mh2sql.cxx --- import existing midas history files (*.hst) into an odbc
sql database.

This new code is enabled by the HAVE_ODBC gunk in the Makefile. If compilation
bombs, please let me know and as a work around, comment out all instances of
HAVE_ODBC from your Makefile.

Limitations:
- mhttpd support for reading history data from odbc sql database is missing
- many sql functions are implemented in a very minimalistic form (i.e. when
defining a history event, we blindly ask sql to create the tables, even if they
already exist - this works, but spams the midas log with sql errors).
- error handling is incomplete: after any sql error, the odbc connection is closed.
- only MySQL (and ascii output) are supported: we use mysql-specific data types
as they match midas types exactly. Code to support PgSQL is present and it used
to work, but is commented out. (At TRIUMF/T2K, we intend to use MySQL exclusively).
- ODBC ascii interface is used, instead of the potentially more efficient binary
interface.

To enable:
- create a MySQL database,
- create $HOME/.odbc.ini (see attached example)
- set ODB "/History/PerVariableHistory" to "1" - the new code is intended to be
used with per-variable history. Per-equipment (traditional) history would work,
but will result in suboptimal layout of SQL tables.
- set ODB "/Logger/ODBC_DSN" to the DSN defined in .odc.ini.
- set ODB "/Logger/ODBC_Debug" to non-zero to enable debugging output from the
new code.

To use the "ascii output" mode:
Included is code to write "ascii" sql output into a text file, instead of using
an actual SQL database. To enable it, set "ODBC_DSN" to
"/path/to/some/text/file" and all SQL output will be written to this file. No
actual SQL database required. This mode exists mostly for debugging the SQL syntax.

Despite limitations, the committed code is fully functional - we are presently
using it to record history data from slow controls of T2K detector tests
(voltages, currents, temperatures).

Comments and suggestions on naming and mapping from odb structures to SQL tables
is very much welcome.

K.O.

I have a history that is embedded in a custom page using

<div class="mjshistory" data-group="SampleCryo" data-panel="SampleTemp" data-scale="30m" style="'+size+' position: relative;left: 640px;top: -205px;"></div>

this works fine when I load the page but seems to cause a timeout when reloading (F5) the page. It used to work fine last year but since a midas update this year it does not work. 

When I manually stop the script when firefox reports that it is slowing down the browse I get the following in the console:

Script terminated by timeout at:
binarySearch@http://xxx.psi.ch:8081/mhistory.js:1051:11
MhistoryGraph.prototype.findMinMax@http://xxx.psi.ch:8081/mhistory.js:1583:28
MhistoryGraph.prototype.loadInitialData/<@http://xxx.psi.ch:8081/mhistory.js:780:15

any ideas what may be causing this?

thanks.

Another hint to the problem, the custom page is accessible via 
http://xxx.psi.ch:8081/?cmd=custom&page=SampleCryo&
once loaded the address changes to where A and B change values as time passes (I guess B-A=30m).
http://xxx.psi.ch:8081/?cmd=custom&page=SampleCryo&&A=1614173369&B=1614175169

I have to reproduce the problem. Can you please send me the full link by direct email. As you know, I'm also at PSI.

Stefan

to plot time dependencies of the monitored detector parameters, say, voltages or temperatures,
one needs to convert the coresponging ADC readings into floats.

One could think of two ways of doing that: 

- one can perform the ADC-->T or ADC-->V conversion in the MIDAS frontend, 
  store their [float] values in the data bank, and plot precalculated parameters vs time

- one can also store in the data bank the ADC readings which typically are short's 
  and convert them into floats (V's or T's) at the plotting time 

The first approach doubles the storage space requirements, and I couldn't find the place where 
one would do the conversion, if stored were the 16-bit ADC readings.

I'm sure this issue has been thought about, so what is the "recommended MIDAS way" of performing 
the ADC -> monitored_number conversion when making MIDAS history plots ?

-- many thanks, regards, Pasha

> to plot time dependencies of the monitored detector parameters, say, voltages or temperatures,
> one needs to convert the coresponging ADC readings into floats.
> 
> One could think of two ways of doing that: 
> 
> - one can perform the ADC-->T or ADC-->V conversion in the MIDAS frontend, 
>   store their [float] values in the data bank, and plot precalculated parameters vs time
> 
> - one can also store in the data bank the ADC readings which typically are short's 
>   and convert them into floats (V's or T's) at the plotting time 
> 
> The first approach doubles the storage space requirements, and I couldn't find the place where 
> one would do the conversion, if stored were the 16-bit ADC readings.
> 
> I'm sure this issue has been thought about, so what is the "recommended MIDAS way" of performing 
> the ADC -> monitored_number conversion when making MIDAS history plots ?

Most experiment go with the second method. The front-end program converts all ADC reading into physicsl 
units, i.e. not only Volt, but even Degrees Centigrade or Tesla or whatever. The slow control part of 
midas then puts these number into /Equipment/<name>/Variables as "float", and the history system picks 
them up from there. This way your history is shown in physical units and not ADC count. Actually the 
recommended slow control framework (check the examples direcotory) does not rely on data banks, but 
puts values directly into the ODB. This is typically done faster, like once per second if a value 
changes, rather than slow control events which are generated maybe once per 10 seconds or once per 
minute. Usually the slow control values are only few compared with trigger data, so a factor of two 
there does not really matter. In the MEG experiment, we have like 400 GB of slow control data per year, 
but 400 TB of trigger data per year.

Best,
Stefan

I have a fe, with 2 "equipments" (2 different types of LV supplies).

Equipment/../Setting has a "Names" key, with the actual channel names (ch1, ch2, ...) of the devices.

Equipment/../Variables has channel states, voltage, etc. 

I also write a separate midas bank for each supply.

When I turn the "Log History" on, 2 things happen which cause troubles:
1. It writes the data of both bank to both the /Equipment/(Device1/Device2)/Variables .
2. I have e.g. 4 channels. In the banks I write current and voltages, so 8 numbers. When I turn on the logger I get an "Array size mismatch" between names and the midas bank size.

The only way around this is to disable the history logging in the equipment, and set "virtual" History events?

First, the writing of banks is completely independent of the history system. Banks go to the log file only, 
while the history is only linked to the "Variables" section in the ODB.

Second, it's advisable to group similar equipment into one. Like if you have five power supplies powering
and experiment, you don't want to have five equipments Supply1, Supply2, ..., but only one equipment
"Power Supplies". In the frontend belonging to that equipment, you define a DEVICE_DRIVER list with
one entry for each power supply. If you interact with an mscb device, there are some helper functions
which simplify the definition of the equipment and which I can send you privately. So your device
driver looks a bit like the one attached.

If you cannot do that and absolutely want separate equipments, please post a complete ODB subtree of your
settings, and I can try to reproduce your problem.

Stefan

======================

DEVICE_DRIVER power_driver[] = {
   {"Power Supply 1", mscbdev, 0, NULL, DF_INPUT | DF_MULTITHREAD},
   {"Power Supply 2", mscbdev, 0, NULL, DF_INPUT | DF_MULTITHREAD},
   {"Power Supply 3", mscbdev, 0, NULL, DF_INPUT | DF_MULTITHREAD},
   {""}
};

...

INT frontend_init()
{
   mscb_define("mscbxxx.psi.ch", "Power Supplies", "Power Supply 1", power_driver, 1, 0, "Output 1", 0.1);
   mscb_define("mscbxxx.psi.ch", "Power Supplies", "Power Supply 1", power_driver, 1, 1, "Output 2", 0.1);
   ...
}

/*-- Function to define MSCB variables in a convenient way ---------*/

void mscb_define(const char *submaster, const char *equipment, const char *devname, 
                 DEVICE_DRIVER *driver, int address, unsigned char var_index, 
                 const char *name, double threshold)
{
   int i, dev_index, chn_index, chn_total;
   char str[256];
   float f_threshold;
   HNDLE hDB;

   cm_get_experiment_database(&hDB, NULL);

   if (submaster && submaster[0]) {
      sprintf(str, "/Equipment/%s/Settings/Devices/%s/Device", equipment, devname);
      db_set_value(hDB, 0, str, submaster, 32, 1, TID_STRING);
      sprintf(str, "/Equipment/%s/Settings/Devices/%s/Pwd", equipment, devname);
      db_set_value(hDB, 0, str, "meg", 32, 1, TID_STRING);
   }

   /* find device in device driver */
   for (dev_index=0 ; driver[dev_index].name[0] ; dev_index++)
      if (equal_ustring(driver[dev_index].name, devname))
         break;

   if (!driver[dev_index].name[0]) {
      cm_msg(MERROR, "mscb_define", "Device \"%s\" not present in device driver list", devname);
      return;
   }

   /* count total number of channels */
   for (i=chn_total=0 ; i<=dev_index ; i++)
      if (((driver[dev_index].flags & DF_INPUT) > 0 && (driver[i].flags & DF_INPUT)) ||
          ((driver[dev_index].flags & DF_OUTPUT) > 0 && (driver[i].flags & DF_OUTPUT)))
      chn_total += driver[i].channels;

   chn_index = driver[dev_index].channels;
   sprintf(str, "/Equipment/%s/Settings/Devices/%s/MSCB Address", equipment, devname);
   db_set_value_index(hDB, 0, str, &address, sizeof(int), chn_index, TID_INT, TRUE);
   sprintf(str, "/Equipment/%s/Settings/Devices/%s/MSCB Index", equipment, devname);
   db_set_value_index(hDB, 0, str, &var_index, sizeof(char), chn_index, TID_BYTE, TRUE);

   if (threshold != -1 && (driver[dev_index].flags & DF_INPUT) > 0) {
     sprintf(str, "/Equipment/%s/Settings/Update Threshold", equipment);
     f_threshold = (float) threshold;
     db_set_value_index(hDB, 0, str, &f_threshold, sizeof(float), chn_total, TID_FLOAT, TRUE);
   }

   if (name && name[0]) {
      sprintf(str, "/Equipment/%s/Settings/Names %s", equipment, devname);
      db_set_value_index(hDB, 0, str, name, 32, chn_total, TID_STRING, TRUE);
   }

   /* increment number of channels for this driver */
   driver[dev_index].channels++;
}

I wanted to have a c++ style driver, e.g. a instance of a "PowerSupply" class. This was not compatible with the list of DEVICE_DRIVER structs, with needs a C function entry point with variable arguments. 

Anyways, I attach my odb. I believe the issue stands regardless of the specific design choice here. Setting the History Log flag copies the banks created to the "Variables" of every equipments initialized, leading to a mismatch between the names array, and the variables. Can be solved by not using FE history events, but Virtual, but the flag in the Equipment is confusing.  

Bank creation in readout function:

for(const auto& d: drivers)
{
...
...
  bk_create(pevent,bk_name, TID_FLOAT, (void **)&pdata);
...			
  std::vector<float> voltage = d->GetVoltage();
  std::vector<float> current = d->GetCurrent();
  for(channels)
  {
    *pdata++ = voltage.at(iChannel);
    *pdata++ = current.at(iChannel);
  }
  bk_close(pevent, pdata);
}