Hi Lakshya,
I did some digging around. What you are seeing is the result of the
switching to MAP_POPULATE by default in LTTng-UST 2.12[1] in commit
4d4838b ("Use MAP_POPULATE to reduce pagefault when available").
The purpose of this change is to avoid taking page faults which tracing,
reducing first-event in a page latency.
In the master branch, this feature has been made configurable for users
who don't want to pre-populate the pages and would rather take page
faults while tracing[2].
Here is an example from LTTng master with map populate per possible CPU:
```
export LTTNG_UST_MAP_POPULATE_POLICY=cpu_possible
# Create session, channels, start tracing, and run test app
# top -n1 -b | grep -E '(MiB|COMMAND|lttng)'
MiB Mem : 32768.0 total, 21883.7 free, 1456.0 used, 9428.3
buff/cache
MiB Swap: 0.0 total, 0.0 free, 0.0 used. 31312.0 avail
Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+
COMMAND
301117 debian 20 0 880176 2760 2760 S 0.0 0.0 0:00.04
lttng-sessiond
301118 debian 20 0 43856 1376 1376 S 0.0 0.0 0:00.01
lttng-runas
301133 debian 20 0 718616 263456 263456 S 0.0 0.8 0:00.17
lttng-consumerd
301135 debian 20 0 6996 0 0 S 0.0 0.0 0:00.05
lttng-runas
# cat /proc/$(pgrep lttng-sessiond)/statm
lttng-sessiond: 220044 690 690 345 0 29900 0
# pmap $(pgrep lttng-sessiond) | grep total
total 880176K
# smem -P lttng-sessiond
PID User Command Swap USS PSS
RSS
301118 debian lttng-sessiond --daemonize 0 344 881
2236
301117 debian lttng-sessiond --daemonize 0 5676 6683
9276
301201 debian /usr/bin/python3 /usr/bin/s 0 8636 10086
12936
# /proc/PID/statm for lttng-consumerd
lttng-consumerd: 1749 0 0 129 0 130 0
# pmap lttng-consumerd-pid | grep total
total kB 6996 1700 472
# smem -P lttng-consumerd
PID User Command Swap USS PSS
RSS
301135 debian lttng-consumerd -u --consu 0 280 563
1700
301211 debian /usr/bin/python3 /usr/bin/s 0 10048 11501
14404
301133 debian lttng-consumerd -u --consu 0 262376 263177
265480
# smem -m | grep -i ust
/dev/shm/lttng-ust-wait-8-1000 1 4 4
/dev/shm/shm-ust-consumer-301133 1 260756 260756
```
When using the none policy:
```
# export LTTNG_UST_MAP_POPULATE_POLICY=none
# as above
Running test app UID 0
procs -----------memory---------- ---swap-- -----io---- -system--
------cpu-----
r b swpd free buff cache si so bi bo in cs us sy
id wa st
1 0 0 21875 0 9434 0 0 39 636 1105 2496 0 1
99 0 0
MiB Mem : 32768.0 total, 21875.0 free, 1458.2 used, 9434.7
buff/cache
MiB Swap: 0.0 total, 0.0 free, 0.0 used. 31309.8 avail
Mem
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+
COMMAND
301616 debian 20 0 880176 2756 2756 S 0.0 0.0 0:00.04
lttng-sessiond
301617 debian 20 0 43856 1392 1392 S 0.0 0.0 0:00.01
lttng-runas
301632 debian 20 0 718612 5416 5416 S 0.0 0.0 0:00.17
lttng-consumerd
301634 debian 20 0 6992 0 0 S 0.0 0.0 0:00.05
lttng-runas
lttng-sessiond: 220044 689 689 345 0 29900 0
total 880176K
PID User Command Swap USS PSS
RSS
301617 debian lttng-sessiond --daemonize 0 344 862
2188
301616 debian lttng-sessiond --daemonize 0 5784 6759
9328
301700 debian /usr/bin/python3 /usr/bin/s 0 8632 10079
12928
lttng-consumerd: 1748 0 0 129 0 129 0
total kB 6992 1580 468
PID User Command Swap USS PSS
RSS
301634 debian lttng-consumerd -u --consu 0 276 536
1580
301632 debian lttng-consumerd -u --consu 0 5672 6433
8652
301710 debian /usr/bin/python3 /usr/bin/s 0 9996 11449
14328
/dev/shm/lttng-ust-wait-8-1000 1 4 4
/dev/shm/shm-ust-consumer-301632 1 4048 4048
```
thanks,
kienan
[1]:
https://github.com/lttng/lttng-ust/commit/4d4838bad480d48424bddc686f5ad0089e28ac94
[2]:
https://github.com/lttng/lttng-ust/commit/97572c0438845cee953ebd3e39615f78bfa405a7
On 3/17/25 2:29 AM, Gour DEV wrote:
Hi, Kienan
Sorry for the late reply.
Looks like in buster the memory is allocated by lttng-consumerd reserved
I buster, the rss is less than the VIRT
root@localhost:~# COLUMNS=500 top -b -n 1 | grep lttng
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+
COMMAND
4095 root 20 0 1003188 31256 4660 S 0.0 0.1 0:03.81
lttng-sessiond
4096 root 20 0 44260 796 0 S 0.0 0.0 0:00.01
lttng-runas
4440 root 20 0 5236020 10224 8756 S 0.0 0.0 2:56.25
lttng-consumerd -- here the VIRT is much more higher than RSS
4443 root 20 0 48048 540 0 S 0.0 0.0 0:00.12
lttng-runas
In bookworm the VIRT and RES are nearly the same only.
root@edgecore-40XKE-j2-101-32:~# COLUMNS=500 top -b -n 1 | grep lttng
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+
COMMAND
4382 root 20 0 1098824 42600 8436 S 0.0 0.1 0:08.87
lttng-sessiond
4403 root 20 0 48928 2116 996 S 0.0 0.0 0:00.00
lttng-runas
5171 root 20 0 9879764 8.9g 8.9g S 0.0 28.7 108:23.53
lttng-consumerd -- here the VRIT is nearly equal to RSS
5173 root 20 0 3680 1028 680 S 0.0 0.0 0:00.88
lttng-runas
Looks like lttng consumerd is allocating and reserving those pages, when
any instrumented application starts.
I am attaching the lttng status output in the mail, please do tell me if
you need any more information regarding this.
These is how we used to create the lttng channels and enable event which
is
same in both buster and bookworm, (number of channels might differ)
def enable_channel(channels, session, subbuf_size, subbuf_num):
for c in channels:
call(['lttng', 'enable-channel', '-u', c, '-s', session, '--subbuf-size',
str(subbuf_size), '--num-subbuf', str(subbuf_num),],
stdout=devnull, stderr=subprocess.STDOUT)
def enable_events(traces, session):
for t in traces:
if 'log-level-only' in t:
log_opt = '--loglevel-only=' + t['log-level-only']
elif 'log-level' in t:
log_opt = '--loglevel=' + t['log-level']
else:
log_opt = ''
else:
call(['lttng', 'enable-event', '-u', t['name'], '-c', t['channel'],
'-s', session], stdout=devnull, stderr=subprocess.STDOUT)
Thank You.
Lakshya
On Wed, Mar 12, 2025 at 8:06 PM <lttng-dev-requ...@lists.lttng.org>
wrote:
Send lttng-dev mailing list submissions to
lttng-dev@lists.lttng.org
To subscribe or unsubscribe via the World Wide Web, visit
https://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev
or, via email, send a message with subject or body 'help' to
lttng-dev-requ...@lists.lttng.org
You can reach the person managing the list at
lttng-dev-ow...@lists.lttng.org
When replying, please edit your Subject line so it is more specific
than "Re: Contents of lttng-dev digest..."
Today's Topics:
1. Re: Memory Consumption High After Upgrading to 2.13 from 2.10
(Kienan Stewart)
2. Re: Memory Consumption High After Upgrading to 2.13 from 2.10
(Gour DEV)
3. Re: Memory Consumption High After Upgrading to 2.13 from 2.10
(Kienan Stewart)
----------------------------------------------------------------------
Message: 1
Date: Tue, 11 Mar 2025 14:55:21 -0400
From: Kienan Stewart <kstew...@efficios.com>
To: Gour DEV <lakshyagou...@gmail.com>, lttng-dev@lists.lttng.org
Subject: Re: Memory Consumption High After Upgrading to 2.13 from 2.10
Message-ID: <38dab5ef-f106-4e57-9e36-b4b30015c...@efficios.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Hi Lakshya,
On 3/11/25 12:25 PM, Gour DEV wrote:
> Hi, Kienan
>
> here is the requested output
>
> root@localhost:~# top -b -n 1 | grep lttng
> 4841 root 20 0 11.5g 11.0g 11.0g S 5.9 35.4
8:39.93
> lttng-c+
> 4824 root 20 0 1098824 26456 5380 S 0.0 0.1
0:07.25
> lttng-s+
> 4825 root 20 0 48872 2188 1012 S 0.0 0.0
0:00.00
> lttng-r+
> 4843 root 20 0 3680 1160 816 S 0.0 0.0
0:00.23
This top output for `localhost` seems very different than the output for
`localhost` in your previous message.
> lttng-r+
> root@localhost:~# nrpco
> bash: nrpco: command not found
> root@localhost:~# nproc
> 16
> root@localhost:~# cat /sys/devices/system/cpu/possible
> 0-15
>
You indicated the bookworm machine has 32 cores, this is showing 16. If
you're comparing a 16 core machine to a 32 core machine, it is very
normal that the memory usage is higher on the 32 core machine.
>
> Most of the process are running as asorcs user but some are running
as root.
So you have two users with instrumented applications.
Given the discrepancies in the information provided I'm finding it a bit
hard to understand what you're looking at.
In general, a channel's shared memory footprint can be estimated
with[1]:
(nSubbuf * subbufSize) * (nCPUs + 1 iff snapshot mode is enabled) *
(nUIDs or nPIDs)
Note that the sub-buffer sizes you are using get rounded to the nearest
larger power of 2. See [2].
thanks,
kienan
[1]: https://lttng.org/docs/v2.13/#doc-channel-buffering-schemes
[2]:
https://lttng.org/man/1/lttng-enable-channel/v2.13/#doc-opt--subbuf-size
------------------------------
Message: 2
Date: Wed, 12 Mar 2025 14:49:07 +0530
From: Gour DEV <lakshyagou...@gmail.com>
To: Kienan Stewart <kstew...@efficios.com>
Cc: lttng-dev@lists.lttng.org
Subject: Re: Memory Consumption High After Upgrading to 2.13 from 2.10
Message-ID:
<CAE9Jrzg7qsabhPiO-0=B1DY3bVo-3FYu2tiJR2Bmb=
nqohn...@mail.gmail.com>
Content-Type: text/plain; charset="utf-8"
Hi, Kienan
I am attaching an screen recording of the behaviour I am seeing in this
mail. The behaviour is same irrespective of the device i use, sorry for
miscommunication in the npocs output (I assumed it was 32), but other
than
that all outputs are same (except the hostname as there are multiple
devices with same lttng config but this memory cosumption is seen on all
the devices).
I had few question
1. Does lltng allocated all the memory it needs and mark it as dirty in
ram
when any process which links/uses lttng-ust runs? (here i tried with one
process but it is same for any of my process)
2. (nSubbuf * subbufSize) * (nCPUs + 1 iff snapshot mode is enabled) *
(nUIDs or nPIDs)
How do we calculate uid in the system is it all uids in the system? is
it
equal to `cat /etc/passwd | wc -l` ?
I will put my calculations according to the above estimate based on all
the
channel i am creating
(4194304*4 + 262144*4 + 16384*4) * (16) * (30 if number user are equal
to
`cat /etc/passwd | wc -l`)B = 7.998046875 GB approx [this is based on
the
start_lttng.py please do correct me if am wrong here.]
But since there are only two users which uses lttng i think the correct
estimate would be
(4194304*4 + 262144*4 + 16384*4) * (16) * (2)B = 546MB
Please do correct me If I am wrong calculations here.
Now, there are a few things here, according to my output lttng is using
11G
which is much more higher than the what is configured.
I am attaching the lttng status and the file which is uses to create the
lttng sessions.
Thank You.
https://drive.google.com/file/d/1tS_ZWEsXDpHZXfWzZHXmWcT0igiIOIaa/view?usp=sharing
-- recording of the behaviour which is seen
https://drive.google.com/file/d/1PrU31oyEw1n9tKETlUtmNGO50s6ywx7p/view?usp=sharing
-- the file which is used to create lttng sessions
On Wed, Mar 12, 2025 at 12:25?AM Kienan Stewart <kstew...@efficios.com>
wrote:
Hi Lakshya,
On 3/11/25 12:25 PM, Gour DEV wrote:
> Hi, Kienan
>
> here is the requested output
>
> root@localhost:~# top -b -n 1 | grep lttng
> 4841 root 20 0 11.5g 11.0g 11.0g S 5.9 35.4
8:39.93
> lttng-c+
> 4824 root 20 0 1098824 26456 5380 S 0.0 0.1
0:07.25
> lttng-s+
> 4825 root 20 0 48872 2188 1012 S 0.0 0.0
0:00.00
> lttng-r+
> 4843 root 20 0 3680 1160 816 S 0.0 0.0
0:00.23
This top output for `localhost` seems very different than the output
for
`localhost` in your previous message.
> lttng-r+
> root@localhost:~# nrpco
> bash: nrpco: command not found
> root@localhost:~# nproc
> 16
> root@localhost:~# cat /sys/devices/system/cpu/possible
> 0-15
>
You indicated the bookworm machine has 32 cores, this is showing 16. If
you're comparing a 16 core machine to a 32 core machine, it is very
normal that the memory usage is higher on the 32 core machine.
>
> Most of the process are running as asorcs user but some are running
as root.
So you have two users with instrumented applications.
Given the discrepancies in the information provided I'm finding it a
bit
hard to understand what you're looking at.
In general, a channel's shared memory footprint can be estimated
with[1]:
(nSubbuf * subbufSize) * (nCPUs + 1 iff snapshot mode is enabled) *
(nUIDs or nPIDs)
Note that the sub-buffer sizes you are using get rounded to the nearest
larger power of 2. See [2].
thanks,
kienan
[1]: https://lttng.org/docs/v2.13/#doc-channel-buffering-schemes
[2]:
https://lttng.org/man/1/lttng-enable-channel/v2.13/#doc-opt--subbuf-size
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <
https://lists.lttng.org/pipermail/lttng-dev/attachments/20250312/57f240d8/attachment-0001.htm
------------------------------
Message: 3
Date: Wed, 12 Mar 2025 10:36:28 -0400
From: Kienan Stewart <kstew...@efficios.com>
To: Gour DEV <lakshyagou...@gmail.com>, lttng-dev@lists.lttng.org
Subject: Re: Memory Consumption High After Upgrading to 2.13 from 2.10
Message-ID: <0f819583-ea8e-468e-9102-e1410d886...@efficios.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Hi Lakshya,
On 3/12/25 5:03 AM, Gour DEV wrote:
Hi, Kienan
I am attaching an screen recording of the behaviour I am seeing in this
mail. The behaviour is same irrespective of the device i use, sorry for
miscommunication in the npocs output (I assumed it was 32), but other
than
that all outputs are same (except the hostname as there are multiple
devices with same lttng config but this memory cosumption is seen on
all
the devices).
I had few question
1. Does lltng allocated all the memory it needs and mark it as dirty in
ram
when any process which links/uses lttng-ust runs? (here i tried with
one
process but it is same for any of my process)
I believe the shared memory for per-CPU data structures is allocated
when an instrumented application connects. There is no pre-allocation
for each possible UID on the system.
You can run your instrumented applications with `LTTNG_UST_DEBUG=1` to
see when the connection happens[1].
2. (nSubbuf * subbufSize) * (nCPUs + 1 iff snapshot mode is enabled) *
(nUIDs or nPIDs)
How do we calculate uid in the system is it all uids in the system? is
it
equal to `cat /etc/passwd | wc -l` ?
nUIDs is the number of distinct UIDs running instrumented applications.
I will put my calculations according to the above estimate based on all
the
channel i am creating
(4194304*4 + 262144*4 + 16384*4) * (16) * (30 if number user are equal
to
`cat /etc/passwd | wc -l`)B = 7.998046875 GB approx [this is based on
the
start_lttng.py please do correct me if am wrong here.]
But since there are only two users which uses lttng i think the correct
estimate would be
(4194304*4 + 262144*4 + 16384*4) * (16) * (2)B = 546MB
The estimate I gave is per-channel.
small channel: (0.015625 MiB * 4) * (16 + 1) = 1.0625 MiB per-channel
per-UID
medium channel: (0.250 MiB * 4) * (16 + 1) = 17.0 MiB per-channel
per-UID
large channel: (4 MiB * 4) * (16 + 1) = 27 2MiB per-channel per-UID
Now, you said you have 0 small channels, 6 medium channels, and 16 large
channels in your session. (Note: I see your script differs from these
stated channel counts).
small: 0 * 1.0625 MiB = 0 MiB per-UID
medium: 6 * 17 MiB = 102 MiB per-UID
large: 16 * 272 MiB = 4352 MiB per-UID
And if you're running instrumented applications with 2 users:
small: 0 MiB * 2 = 0 MiB with 2 UIDs
medium: 102 MiB * 2 = 204 MiB with 2 UIDs
large: 4352 MiB * 2 = 8704 MiB with 2 UIDs
Now this is just an estimation for the per-CPU ring buffers only, and
you numbers aren't hugely off so without analyzing your specific system
it doesn't seem to be that strange to me.
If I take the number of channels I see in your script, it becomes:
small: 0 MiB with 2 UIDs
medium: 136 MiB with 2 UIDs
large: 7616 MiB with 2 UIDs
total: 7.57 GiB with 2 UIDs
Please do correct me If I am wrong calculations here.
Now, there are a few things here, according to my output lttng is using
11G
which is much more higher than the what is configured.
I have no idea what 'service start spyder' is doing. Maybe it's running
instrumented applications with an extra user that you didn't expect? I
can't help you with that aspect of your system.
The above estimated 7.57 GiB with 2 UIDs would be 11.35 GiB with 3 UIDs
so maybe?
I'd recommend you read your verbose sessiond log so see which
applications are connecting and with which UIDs.
I am attaching the lttng status and the file which is uses to create
the
lttng sessions.
Thank You.
In any case, the information you have given to date hasn't demonstrated
to me in a tangible manner that you are seeing a difference related to
the version of LTTng being used.
thanks,
kienan
[1]:
https://lttng.org/man/3/lttng-ust/v2.13/#doc-_environment_variables
------------------------------
Subject: Digest Footer
_______________________________________________
lttng-dev mailing list
lttng-dev@lists.lttng.org
https://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev
------------------------------
End of lttng-dev Digest, Vol 203, Issue 7
*****************************************
