qdisc running
Subsystems (2)
Jesper,
You asked at the meeting the point to qdisc running. Original intent is to allow only one cpu to enter the lower half of the qdisc path. IOW, if one cpu was already in the qdisc then that guy could be used to dequeue packets. i.e this is good for batching. Original idea was Herbert's with major improvement from Eric and a small one from me.
For history of different tried approaches look at:
Look at slide 2: http://vger.kernel.org/netconf2011_slides/jamal_netconf2011.pdf
then download the **amazing** flash animations which describe
that history.
http://vger.kernel.org/netconf2011_slides/netconf-2011-flash.tgz
Follow the bullets in slide2 and map to the flash animations.
If you go over them, you'll see it is still needed.
I think someone oughta put those **amazing** animations on some website;->
cheers,
jamal
Quoted 3 lines · On Sun, 19 Oct 2014 15:24:42 -0400 Jamal Hadi Salim <[email protected]> wrote:
Jesper,
You asked at the meeting the point to qdisc running.
Talking about __QDISC___STATE_RUNNING see slide 9/16:
http://people.netfilter.org/hawk/presentations/LinuxPlumbers2014/performance_tx_qdisc_bulk_LPC2014.pdf
Quoted 5 lines
Original intent is to allow only one cpu to enter the lower half of the qdisc path. IOW, if one cpu was already in the qdisc then that guy could be used to dequeue packets. i.e this is good for batching. Original idea was Herbert's with major improvement from Eric and a small one from me.
I guess it is good for our recent dequeue batching. But I think/hope we can come up with a scheme that does not requires 6 lock/unlock operations (as illustrated on slide 9).
John and I have talked about doing a lockless qdisc, but maintaining this __QDISC___STATE_RUNNING in a lockless scenario, would cost us extra atomic ops...
Are we still sure, that this model of only allowing a single CPU in the dequeue path, is still the best solution? (The TXQ lock should already protect several CPUs in this code path).
Quoted 3 lines
For history of different tried approaches look at:
Look at slide 2:
http://vger.kernel.org/netconf2011_slides/jamal_netconf2011.pdf
I can see that you really needed the budget/fairness in the dequeue loop, that we recently mangled with.
Quoted 5 lines
then download the **amazing** flash animations which describe that history.
http://vger.kernel.org/netconf2011_slides/netconf-2011-flash.tgzFollow the bullets in slide2 and map to the flash animations.
What tool do I use to play these SWF files? (I tried VLC but no luck).
Quoted 1 line
If you go over them, you'll see it is still needed.
Too bad, I would like to avoid the second
Quoted 2 lines
I think someone oughta put those **amazing** animations on some website;->
I hope someone else can pick that up ;-)
Signature · 5 lines
Best regards, Jesper Dangaard Brouer MSc.CS, Sr. Network Kernel Developer at Red Hat Author of http://www.iptv-analyzer.org LinkedIn: http://www.linkedin.com/in/brouer
Quoted 5 lines · Jesper Dangaard Brouer
On Sun, 19 Oct 2014 15:24:42 -0400 Jamal Hadi Salim <[email protected]> wrote:
I guess it is good for our recent dequeue batching.
It is i think ;->
Quoted 4 lines
But I think/hope we
can come up with a scheme that does not requires 6 lock/unlock operations (as illustrated on slide 9).
To be clear:
2 locks + 2 unlock and 2 atomic ops.
Quoted 4 lines
John and I have talked about doing a lockless qdisc, but maintaining this
__QDISC___STATE_RUNNINGin a lockless scenario, would cost us extra atomic ops...
In the animation this __QDISC___STATE_RUNNING is shown as "occupied" flag. It is like someone is in the toilet and you cant come in;-> They have to finish dropping the packages into the toilet^Whardware ;-> If it is occupied, you put your package outside and go.
Quoted 2 lines
Are we still sure, that this model of only allowing a single CPU in the dequeue path, is still the best solution?
For sure it is the best if you want to batch. Look at that last orange guy picking all the packages (busylock.swf). This is where all the batching would happen.
Quoted 2 lines
(The TXQ lock should already
protect several CPUs in this code path).
Note:
Maybe for the orange guy (the dequeur) the tx lock could be avoided? Double check the code. Important to note under busy period contention is reduced to :
1 lock + 1 unlock + 2 atomic ops for N-1 CPUs. The orange guy on the other hand is doing 2 lock/unlock.
Quoted 3 lines
I can see that you really needed the budget/fairness in the dequeue loop, that we recently mangled with.
Yes, fairness is needed so the orange guy doesnt spend all his cycles doing all the work (that was the basis of my presentation); unless that is not an issue and the scheduler would move things away from that cpu.
Quoted 2 lines
What tool do I use to play these SWF files? (I tried VLC but no luck).
Firefox should work fine.
cheers,
jamal