* [LibreQoS] cpumap-pping 1.0 final
@ 2022-11-15 14:08 Herbert Wolverson
2022-11-15 14:44 ` Dave Taht
0 siblings, 1 reply; 2+ messages in thread
From: Herbert Wolverson @ 2022-11-15 14:08 UTC (permalink / raw)
To: libreqos
[-- Attachment #1: Type: text/plain, Size: 1660 bytes --]
Hey!
Just wanted to announce that cpumap-pping is now tagged as the 1.0 release.
The one major change from RC1 is that the xdp_pping utility no longer does
any cleaning - it just reports statistics. That makes its execution a LOT
faster, and removes a stalling issue that was showing up under heavy (5
gbit/s+) load.
Under the hood, each "report flow" now carries a "recycle time" and a
"fresh data" flag. They are also stored in an "LRU map" rather than a
strict hash map - so if the buffer fills up, it automatically (in kernel
space) evicts (replaces) the oldest member rather than pausing to report
the event. This in turn allows for all memory allocation to be up-front (on
load), so there are *no* memory allocations or reallocations within the
running RTT tracking system.
Once a "report flow" has achieved its target number of samples, it stops
collecting data. 30 seconds (by default, it's configurable) later, the flow
is "recycled" - its buffer is empties, "fresh data" is set to false, and
the recycle time bumped to "now + 30 seconds". If data arrives, "fresh
data" is set to true - and the reporting tool will show the new data. If no
data arrives, the flow ages slowly to the bottom of the list and is
replaced with no memory allocation/deallocation/reallocation.
This has largely eliminated the occasional variance in execution time (it
never spiked above 6000 nanoseconds to begin with; but now it's
consistently under 0.0005 ms added ping time; you won't notice it). More
importantly, it completely eradicated the high-load behavior of the
xdp_pping sitting and spinning for long periods of times under high lock
contention.
-- Herbert
[-- Attachment #2: Type: text/html, Size: 1948 bytes --]
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [LibreQoS] cpumap-pping 1.0 final
2022-11-15 14:08 [LibreQoS] cpumap-pping 1.0 final Herbert Wolverson
@ 2022-11-15 14:44 ` Dave Taht
0 siblings, 0 replies; 2+ messages in thread
From: Dave Taht @ 2022-11-15 14:44 UTC (permalink / raw)
To: Herbert Wolverson; +Cc: libreqos
yay!
https://twitter.com/mtrantalainen/status/1592525968861990914
On Tue, Nov 15, 2022 at 6:08 AM Herbert Wolverson via LibreQoS
<libreqos@lists.bufferbloat.net> wrote:
>
> Hey!
>
> Just wanted to announce that cpumap-pping is now tagged as the 1.0 release.
>
> The one major change from RC1 is that the xdp_pping utility no longer does any cleaning - it just reports statistics. That makes its execution a LOT faster, and removes a stalling issue that was showing up under heavy (5 gbit/s+) load.
>
> Under the hood, each "report flow" now carries a "recycle time" and a "fresh data" flag. They are also stored in an "LRU map" rather than a strict hash map - so if the buffer fills up, it automatically (in kernel space) evicts (replaces) the oldest member rather than pausing to report the event. This in turn allows for all memory allocation to be up-front (on load), so there are no memory allocations or reallocations within the running RTT tracking system.
>
> Once a "report flow" has achieved its target number of samples, it stops collecting data. 30 seconds (by default, it's configurable) later, the flow is "recycled" - its buffer is empties, "fresh data" is set to false, and the recycle time bumped to "now + 30 seconds". If data arrives, "fresh data" is set to true - and the reporting tool will show the new data. If no data arrives, the flow ages slowly to the bottom of the list and is replaced with no memory allocation/deallocation/reallocation.
>
> This has largely eliminated the occasional variance in execution time (it never spiked above 6000 nanoseconds to begin with; but now it's consistently under 0.0005 ms added ping time; you won't notice it). More importantly, it completely eradicated the high-load behavior of the xdp_pping sitting and spinning for long periods of times under high lock contention.
>
> -- Herbert
> _______________________________________________
> LibreQoS mailing list
> LibreQoS@lists.bufferbloat.net
> https://lists.bufferbloat.net/listinfo/libreqos
--
This song goes out to all the folk that thought Stadia would work:
https://www.linkedin.com/posts/dtaht_the-mushroom-song-activity-6981366665607352320-FXtz
Dave Täht CEO, TekLibre, LLC
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2022-11-15 14:44 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-11-15 14:08 [LibreQoS] cpumap-pping 1.0 final Herbert Wolverson
2022-11-15 14:44 ` Dave Taht
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox