vhost-kernel net on pasta: from 26 to 37Gbit/s

newer
[PATCH] fedora: Hide restorecon(8)...

Eugenio Perez Martin

20 May 2025 20 May '25

5:09 p.m.

Hi! Some updates on the integration. The main culprit was to allow pasta to keep reading packets using the regular read() on the tap device. I thought that part was completely disabled, but I guess the kernel is able to omit the notification on tap as long as the userspace does not read it. My scenario: in different CPUs, all in the same NUMA. I run iperf server on CPU 11 with "iperf3 -A 11 -s". All odds CPUs are isolated with isolcpus=1,3,... nohz=on nohz_full=1,3,... With vanilla pasta isolated to CPUs 1,3 with taskset, and just --config-net option, and running iperf with "iperf3 -A 5 -c 10.6.68.254 -w 8M": - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 30.7 GBytes 26.4 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 30.7 GBytes 26.3 Gbits/sec receiver Now trying with the vhost patches we get a slightly worse performance: - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 25.5 GBytes 21.9 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 25.5 GBytes 21.8 Gbits/sec receiver Now vhost patch still lacks optimizations like disabling notifications or batch more rx available buffer notifications. At the moment it refills the rx buffers in each iteration, and does not set the no_notify bit which makes the kernel skip the used buffer notifications if pasta is actively checking the queue, which is not optimal. Now if I isolate the vhost kernel thread [1] I get way more performance as expected: - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 43.1 GBytes 37.1 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 43.1 GBytes 36.9 Gbits/sec receiver After analyzing perf output, rep_movs_alternative is the most called function in the three iperf3 (~20%Self), passt.avx2 (~15%Self) and vhost (~15%Self), But I don't see any of them consuming 100% of CPU in top: pasta consumes ~85% %CPU, both iperf3 client and server consumes 60%, and vhost consumes ~53%. So... I have mixed feelings about this :). By "default" it seems to have less performance, but my test is maybe too synthetic. There is room for improvement with the mentioned optimizations so I'd continue applying them, continuing with UDP and TCP zerocopy, and developing zerocopy vhost rx. With these numbers I think the series should not be merged at the moment. I could send it as RFC if you want but I've not applied the comments the first one received, POC style :). Thanks! [1] Notes to reproduce it, I'm able to see it with top -H and then set with taskset. Either the latest changes on the module or the way pasta behaves does not allow me to see in classical ps output.

Show replies by date

Jason Wang

21 May 21 May

2:57 a.m.

On Tue, May 20, 2025 at 11:10 PM Eugenio Perez Martin wrote:

...

Hi!

Some updates on the integration. The main culprit was to allow pasta to keep reading packets using the regular read() on the tap device. I thought that part was completely disabled, but I guess the kernel is able to omit the notification on tap as long as the userspace does not read it.

My scenario: in different CPUs, all in the same NUMA. I run iperf server on CPU 11 with "iperf3 -A 11 -s". All odds CPUs are isolated with isolcpus=1,3,... nohz=on nohz_full=1,3,...

With vanilla pasta isolated to CPUs 1,3 with taskset, and just --config-net option, and running iperf with "iperf3 -A 5 -c 10.6.68.254 -w 8M": - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 30.7 GBytes 26.4 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 30.7 GBytes 26.3 Gbits/sec receiver

Now trying with the vhost patches we get a slightly worse performance: - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 25.5 GBytes 21.9 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 25.5 GBytes 21.8 Gbits/sec receiver

Now vhost patch still lacks optimizations like disabling notifications or batch more rx available buffer notifications. At the moment it refills the rx buffers in each iteration, and does not set the no_notify bit which makes the kernel skip the used buffer notifications if pasta is actively checking the queue, which is not optimal.

Now if I isolate the vhost kernel thread [1] I get way more performance as expected: - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 43.1 GBytes 37.1 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 43.1 GBytes 36.9 Gbits/sec receiver

After analyzing perf output, rep_movs_alternative is the most called function in the three iperf3 (~20%Self), passt.avx2 (~15%Self) and vhost (~15%Self), But I don't see any of them consuming 100% of CPU in top: pasta consumes ~85% %CPU, both iperf3 client and server consumes 60%, and vhost consumes ~53%.

So... I have mixed feelings about this :). By "default" it seems to have less performance, but my test is maybe too synthetic. There is room for improvement with the mentioned optimizations so I'd continue applying them, continuing with UDP and TCP zerocopy, and developing zerocopy vhost rx. With these numbers I think the series should not be merged at the moment. I could send it as RFC if you want but I've not applied the comments the first one received, POC style :).

Have you pinned pasta in a specific CPU? Note that vhost will inherit the affinity so there could be some contention if you do that. Thanks

...

Thanks!

[1] Notes to reproduce it, I'm able to see it with top -H and then set with taskset. Either the latest changes on the module or the way pasta behaves does not allow me to see in classical ps output.

Eugenio Perez Martin

7:37 a.m.

On Wed, May 21, 2025 at 2:57 AM Jason Wang wrote:

...

On Tue, May 20, 2025 at 11:10 PM Eugenio Perez Martin wrote:

...
Hi!

Some updates on the integration. The main culprit was to allow pasta to keep reading packets using the regular read() on the tap device. I thought that part was completely disabled, but I guess the kernel is able to omit the notification on tap as long as the userspace does not read it.

My scenario: in different CPUs, all in the same NUMA. I run iperf server on CPU 11 with "iperf3 -A 11 -s". All odds CPUs are isolated with isolcpus=1,3,... nohz=on nohz_full=1,3,...

With vanilla pasta isolated to CPUs 1,3 with taskset, and just --config-net option, and running iperf with "iperf3 -A 5 -c 10.6.68.254 -w 8M": - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 30.7 GBytes 26.4 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 30.7 GBytes 26.3 Gbits/sec receiver

Now trying with the vhost patches we get a slightly worse performance: - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 25.5 GBytes 21.9 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 25.5 GBytes 21.8 Gbits/sec receiver

Now vhost patch still lacks optimizations like disabling notifications or batch more rx available buffer notifications. At the moment it refills the rx buffers in each iteration, and does not set the no_notify bit which makes the kernel skip the used buffer notifications if pasta is actively checking the queue, which is not optimal.

Now if I isolate the vhost kernel thread [1] I get way more performance as expected: - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 43.1 GBytes 37.1 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 43.1 GBytes 36.9 Gbits/sec receiver

After analyzing perf output, rep_movs_alternative is the most called function in the three iperf3 (~20%Self), passt.avx2 (~15%Self) and vhost (~15%Self), But I don't see any of them consuming 100% of CPU in top: pasta consumes ~85% %CPU, both iperf3 client and server consumes 60%, and vhost consumes ~53%.

So... I have mixed feelings about this :). By "default" it seems to have less performance, but my test is maybe too synthetic. There is room for improvement with the mentioned optimizations so I'd continue applying them, continuing with UDP and TCP zerocopy, and developing zerocopy vhost rx. With these numbers I think the series should not be merged at the moment. I could send it as RFC if you want but I've not applied the comments the first one received, POC style :).

Have you pinned pasta in a specific CPU? Note that vhost will inherit the affinity so there could be some contention if you do that.

Yes, pasta was pinned to 1,3 and vhost to 7.

Stefano Brivio

12:08 p.m.

On Tue, 20 May 2025 17:09:44 +0200 Eugenio Perez Martin wrote:

...

[...]

Now if I isolate the vhost kernel thread [1] I get way more performance as expected: - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 43.1 GBytes 37.1 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 43.1 GBytes 36.9 Gbits/sec receiver

After analyzing perf output, rep_movs_alternative is the most called function in the three iperf3 (~20%Self), passt.avx2 (~15%Self) and vhost (~15%Self)

Interesting... s/most called function/function using the most cycles/, I suppose. So it looks somewhat similar to https://archives.passt.top/passt-dev/20241017021027.2ac9ea53@elisabeth/ now?

...

But I don't see any of them consuming 100% of CPU in top: pasta consumes ~85% %CPU, both iperf3 client and server consumes 60%, and vhost consumes ~53%.

So... I have mixed feelings about this :). By "default" it seems to have less performance, but my test is maybe too synthetic.

Well, surely we can't ask Podman users to pin specific stuff to given CPU threads. :)

...

There is room for improvement with the mentioned optimizations so I'd continue applying them, continuing with UDP and TCP zerocopy, and developing zerocopy vhost rx.

That definitely makes sense to me.

...

With these numbers I think the series should not be merged at the moment. I could send it as RFC if you want but I've not applied the comments the first one received, POC style :).

I don't think it's really needed for you to spend time on semi-polishing something just to have an RFC if you're still working on it. I guess the implementation will change substantially anyway once you factor in further optimisations. -- Stefano

Eugenio Perez Martin

12:35 p.m.

On Wed, May 21, 2025 at 12:09 PM Stefano Brivio wrote:

...

On Tue, 20 May 2025 17:09:44 +0200 Eugenio Perez Martin wrote:

...
[...]

Now if I isolate the vhost kernel thread [1] I get way more performance as expected: - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 43.1 GBytes 37.1 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 43.1 GBytes 36.9 Gbits/sec receiver

After analyzing perf output, rep_movs_alternative is the most called function in the three iperf3 (~20%Self), passt.avx2 (~15%Self) and vhost (~15%Self)

Interesting... s/most called function/function using the most cycles/, I suppose.

Right!

...

So it looks somewhat similar to

https://archives.passt.top/passt-dev/20241017021027.2ac9ea53@elisabeth/

now?

Kind of. Below tcp_sendmsg_locked I don't see sk_page_frag_refill but skb_do_copy_data_nocache. Not sure if that means something, as it should not be affected by vhost.

...

...
But I don't see any of them consuming 100% of CPU in top: pasta consumes ~85% %CPU, both iperf3 client and server consumes 60%, and vhost consumes ~53%.

So... I have mixed feelings about this :). By "default" it seems to have less performance, but my test is maybe too synthetic.

Well, surely we can't ask Podman users to pin specific stuff to given CPU threads. :)

Yes but maybe the result changes under the right schedule? I'm isolating the CPUs entirely, which is not the usual case for pasta for sure :).

...

...
There is room for improvement with the mentioned optimizations so I'd continue applying them, continuing with UDP and TCP zerocopy, and developing zerocopy vhost rx.

That definitely makes sense to me.

Good!

...

...
With these numbers I think the series should not be merged at the moment. I could send it as RFC if you want but I've not applied the comments the first one received, POC style :).

I don't think it's really needed for you to spend time on semi-polishing something just to have an RFC if you're still working on it. I guess the implementation will change substantially anyway once you factor in further optimisations.

Agree! I'll keep iterating on this then. Thanks!

Eugenio Perez Martin

6 Jun 6 Jun

4:32 p.m.

On Wed, May 21, 2025 at 12:35 PM Eugenio Perez Martin wrote:

...

On Wed, May 21, 2025 at 12:09 PM Stefano Brivio wrote:

...
On Tue, 20 May 2025 17:09:44 +0200 Eugenio Perez Martin wrote:

...
[...]

Now if I isolate the vhost kernel thread [1] I get way more performance as expected: - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 43.1 GBytes 37.1 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 43.1 GBytes 36.9 Gbits/sec receiver

After analyzing perf output, rep_movs_alternative is the most called function in the three iperf3 (~20%Self), passt.avx2 (~15%Self) and vhost (~15%Self)

Interesting... s/most called function/function using the most cycles/, I suppose.

Right!

...
So it looks somewhat similar to

https://archives.passt.top/passt-dev/20241017021027.2ac9ea53@elisabeth/

now?

Kind of. Below tcp_sendmsg_locked I don't see sk_page_frag_refill but skb_do_copy_data_nocache. Not sure if that means something, as it should not be affected by vhost.

...
...
But I don't see any of them consuming 100% of CPU in top: pasta consumes ~85% %CPU, both iperf3 client and server consumes 60%, and vhost consumes ~53%.

So... I have mixed feelings about this :). By "default" it seems to have less performance, but my test is maybe too synthetic.

Well, surely we can't ask Podman users to pin specific stuff to given CPU threads. :)

Yes but maybe the result changes under the right schedule? I'm isolating the CPUs entirely, which is not the usual case for pasta for sure :).

...
...
There is room for improvement with the mentioned optimizations so I'd continue applying them, continuing with UDP and TCP zerocopy, and developing zerocopy vhost rx.

That definitely makes sense to me.

Good!

...
...
With these numbers I think the series should not be merged at the moment. I could send it as RFC if you want but I've not applied the comments the first one received, POC style :).

I don't think it's really needed for you to spend time on semi-polishing something just to have an RFC if you're still working on it. I guess the implementation will change substantially anyway once you factor in further optimisations.

Agree! I'll keep iterating on this then.

Actually, if I remove all the taskset etc, and trust the kernel scheduler, vanilla pasta gives me: [pasta@virtlab716 ~]$ /home/passt/pasta --config-net iperf3 -c 10.6.68.254 -w 8M Connecting to host 10.6.68.254, port 5201 [ 5] local 10.6.68.20 port 40408 connected to 10.6.68.254 port 5201 [ ID] Interval Transfer Bitrate Retr Cwnd [ 5] 0.00-1.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 1.00-2.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 2.00-3.00 sec 3.12 GBytes 26.8 Gbits/sec 0 25.4 MBytes [ 5] 3.00-4.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 4.00-5.00 sec 3.10 GBytes 26.6 Gbits/sec 0 25.4 MBytes [ 5] 5.00-6.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 6.00-7.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 7.00-8.00 sec 3.09 GBytes 26.6 Gbits/sec 0 25.4 MBytes [ 5] 8.00-9.00 sec 3.08 GBytes 26.5 Gbits/sec 0 25.4 MBytes [ 5] 9.00-10.00 sec 3.10 GBytes 26.6 Gbits/sec 0 25.4 MBytes - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 31.0 GBytes 26.7 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 31.0 GBytes 26.5 Gbits/sec receiver And with vhost-net : [pasta@virtlab716 ~]$ /home/passt/pasta --config-net iperf3 -c 10.6.68.254 -w 8M ... Connecting to host 10.6.68.254, port 5201 [ 5] local 10.6.68.20 port 46720 connected to 10.6.68.254 port 5201 [ ID] Interval Transfer Bitrate Retr Cwnd [ 5] 0.00-1.00 sec 4.17 GBytes 35.8 Gbits/sec 0 11.9 MBytes [ 5] 1.00-2.00 sec 4.17 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 2.00-3.00 sec 4.16 GBytes 35.7 Gbits/sec 0 11.9 MBytes [ 5] 3.00-4.00 sec 4.14 GBytes 35.6 Gbits/sec 0 11.9 MBytes [ 5] 4.00-5.00 sec 4.16 GBytes 35.7 Gbits/sec 0 11.9 MBytes [ 5] 5.00-6.00 sec 4.16 GBytes 35.8 Gbits/sec 0 11.9 MBytes [ 5] 6.00-7.00 sec 4.18 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 7.00-8.00 sec 4.19 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 8.00-9.00 sec 4.18 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 9.00-10.00 sec 4.18 GBytes 35.9 Gbits/sec 0 11.9 MBytes - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 41.7 GBytes 35.8 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 41.7 GBytes 35.7 Gbits/sec receiver If I go the extra mile and disable notifications (it might be just noise, but...) [pasta@virtlab716 ~]$ /home/passt/pasta --config-net iperf3 -c 10.6.68.254 -w 8M ... Connecting to host 10.6.68.254, port 5201 [ 5] local 10.6.68.20 port 56590 connected to 10.6.68.254 port 5201 [ ID] Interval Transfer Bitrate Retr Cwnd [ 5] 0.00-1.00 sec 4.19 GBytes 36.0 Gbits/sec 0 12.4 MBytes [ 5] 1.00-2.00 sec 4.18 GBytes 35.9 Gbits/sec 0 12.4 MBytes [ 5] 2.00-3.00 sec 4.18 GBytes 35.9 Gbits/sec 0 12.4 MBytes [ 5] 3.00-4.00 sec 4.20 GBytes 36.1 Gbits/sec 0 12.4 MBytes [ 5] 4.00-5.00 sec 4.21 GBytes 36.2 Gbits/sec 0 12.4 MBytes [ 5] 5.00-6.00 sec 4.21 GBytes 36.1 Gbits/sec 0 12.4 MBytes [ 5] 6.00-7.00 sec 4.20 GBytes 36.1 Gbits/sec 0 12.4 MBytes [ 5] 7.00-8.00 sec 4.23 GBytes 36.4 Gbits/sec 0 12.4 MBytes [ 5] 8.00-9.00 sec 4.24 GBytes 36.4 Gbits/sec 0 12.4 MBytes [ 5] 9.00-10.00 sec 4.21 GBytes 36.2 Gbits/sec 0 12.4 MBytes - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 42.1 GBytes 36.1 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 42.1 GBytes 36.0 Gbits/sec receiver So I guess the best is to actually run performance tests closer to real-world workload against the new version and see if it works better? Thanks!

Stefano Brivio

6:37 p.m.

On Fri, 6 Jun 2025 16:32:38 +0200 Eugenio Perez Martin wrote:

...

On Wed, May 21, 2025 at 12:35 PM Eugenio Perez Martin wrote:

...
On Wed, May 21, 2025 at 12:09 PM Stefano Brivio wrote:

...
On Tue, 20 May 2025 17:09:44 +0200 Eugenio Perez Martin wrote:

...
[...]

Now if I isolate the vhost kernel thread [1] I get way more performance as expected: - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 43.1 GBytes 37.1 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 43.1 GBytes 36.9 Gbits/sec receiver

After analyzing perf output, rep_movs_alternative is the most called function in the three iperf3 (~20%Self), passt.avx2 (~15%Self) and vhost (~15%Self)

Interesting... s/most called function/function using the most cycles/, I suppose.

Right!

...
So it looks somewhat similar to

https://archives.passt.top/passt-dev/20241017021027.2ac9ea53@elisabeth/

now?

Kind of. Below tcp_sendmsg_locked I don't see sk_page_frag_refill but skb_do_copy_data_nocache. Not sure if that means something, as it should not be affected by vhost.

...
...
But I don't see any of them consuming 100% of CPU in top: pasta consumes ~85% %CPU, both iperf3 client and server consumes 60%, and vhost consumes ~53%.

So... I have mixed feelings about this :). By "default" it seems to have less performance, but my test is maybe too synthetic.

Well, surely we can't ask Podman users to pin specific stuff to given CPU threads. :)

Yes but maybe the result changes under the right schedule? I'm isolating the CPUs entirely, which is not the usual case for pasta for sure :).

...
...
There is room for improvement with the mentioned optimizations so I'd continue applying them, continuing with UDP and TCP zerocopy, and developing zerocopy vhost rx.

That definitely makes sense to me.

Good!

...
...
With these numbers I think the series should not be merged at the moment. I could send it as RFC if you want but I've not applied the comments the first one received, POC style :).

I don't think it's really needed for you to spend time on semi-polishing something just to have an RFC if you're still working on it. I guess the implementation will change substantially anyway once you factor in further optimisations.

Agree! I'll keep iterating on this then.

Actually, if I remove all the taskset etc, and trust the kernel scheduler, vanilla pasta gives me: [pasta@virtlab716 ~]$ /home/passt/pasta --config-net iperf3 -c 10.6.68.254 -w 8M Connecting to host 10.6.68.254, port 5201 [ 5] local 10.6.68.20 port 40408 connected to 10.6.68.254 port 5201 [ ID] Interval Transfer Bitrate Retr Cwnd [ 5] 0.00-1.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 1.00-2.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 2.00-3.00 sec 3.12 GBytes 26.8 Gbits/sec 0 25.4 MBytes [ 5] 3.00-4.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 4.00-5.00 sec 3.10 GBytes 26.6 Gbits/sec 0 25.4 MBytes [ 5] 5.00-6.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 6.00-7.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 7.00-8.00 sec 3.09 GBytes 26.6 Gbits/sec 0 25.4 MBytes [ 5] 8.00-9.00 sec 3.08 GBytes 26.5 Gbits/sec 0 25.4 MBytes [ 5] 9.00-10.00 sec 3.10 GBytes 26.6 Gbits/sec 0 25.4 MBytes - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 31.0 GBytes 26.7 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 31.0 GBytes 26.5 Gbits/sec receiver

And with vhost-net : [pasta@virtlab716 ~]$ /home/passt/pasta --config-net iperf3 -c 10.6.68.254 -w 8M ... Connecting to host 10.6.68.254, port 5201 [ 5] local 10.6.68.20 port 46720 connected to 10.6.68.254 port 5201 [ ID] Interval Transfer Bitrate Retr Cwnd [ 5] 0.00-1.00 sec 4.17 GBytes 35.8 Gbits/sec 0 11.9 MBytes [ 5] 1.00-2.00 sec 4.17 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 2.00-3.00 sec 4.16 GBytes 35.7 Gbits/sec 0 11.9 MBytes [ 5] 3.00-4.00 sec 4.14 GBytes 35.6 Gbits/sec 0 11.9 MBytes [ 5] 4.00-5.00 sec 4.16 GBytes 35.7 Gbits/sec 0 11.9 MBytes [ 5] 5.00-6.00 sec 4.16 GBytes 35.8 Gbits/sec 0 11.9 MBytes [ 5] 6.00-7.00 sec 4.18 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 7.00-8.00 sec 4.19 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 8.00-9.00 sec 4.18 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 9.00-10.00 sec 4.18 GBytes 35.9 Gbits/sec 0 11.9 MBytes - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 41.7 GBytes 35.8 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 41.7 GBytes 35.7 Gbits/sec receiver

If I go the extra mile and disable notifications (it might be just noise, but...) [pasta@virtlab716 ~]$ /home/passt/pasta --config-net iperf3 -c 10.6.68.254 -w 8M ... Connecting to host 10.6.68.254, port 5201 [ 5] local 10.6.68.20 port 56590 connected to 10.6.68.254 port 5201 [ ID] Interval Transfer Bitrate Retr Cwnd [ 5] 0.00-1.00 sec 4.19 GBytes 36.0 Gbits/sec 0 12.4 MBytes [ 5] 1.00-2.00 sec 4.18 GBytes 35.9 Gbits/sec 0 12.4 MBytes [ 5] 2.00-3.00 sec 4.18 GBytes 35.9 Gbits/sec 0 12.4 MBytes [ 5] 3.00-4.00 sec 4.20 GBytes 36.1 Gbits/sec 0 12.4 MBytes [ 5] 4.00-5.00 sec 4.21 GBytes 36.2 Gbits/sec 0 12.4 MBytes [ 5] 5.00-6.00 sec 4.21 GBytes 36.1 Gbits/sec 0 12.4 MBytes [ 5] 6.00-7.00 sec 4.20 GBytes 36.1 Gbits/sec 0 12.4 MBytes [ 5] 7.00-8.00 sec 4.23 GBytes 36.4 Gbits/sec 0 12.4 MBytes [ 5] 8.00-9.00 sec 4.24 GBytes 36.4 Gbits/sec 0 12.4 MBytes [ 5] 9.00-10.00 sec 4.21 GBytes 36.2 Gbits/sec 0 12.4 MBytes - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 42.1 GBytes 36.1 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 42.1 GBytes 36.0 Gbits/sec receiver

So I guess the best is to actually run performance tests closer to real-world workload against the new version and see if it works better?

Well, that's certainly a possibility. I'd say the biggest value for vhost-net usage in pasta is reaching throughput figures that are comparable with veth, with or without multithreading (keeping an eye on bytes per cycle, of course), with or without kernel changes, so that users won't need to choose between rootless and performance anymore. It would also simplify things in Podman quite a lot (and to some extent in rootlesskit / Docker as well). We're pretty much there with virtual machines, just not quite with containers (which is somewhat ironic, but of course there's a good reason for that). If we're clearly wasting cycles in vhost-net (because of the bounce buffer, plus something else perhaps?) *and* there's a somewhat possible solution for that in sight *and* the interface would change anyway, running throughput tests and polishing up the current version with a half-baked solution at the moment sounds a bit wasteful to me. But if one of those assumptions doesn't hold, or if you feel the need to consolidate the current status, perhaps polishing up the current version right now and actually evaluating throughput (as well as overhead) makes sense to me, yes. -- Stefano

Eugenio Perez Martin

9 Jun 9 Jun

11:59 a.m.

On Fri, Jun 6, 2025 at 6:37 PM Stefano Brivio wrote:

...

On Fri, 6 Jun 2025 16:32:38 +0200 Eugenio Perez Martin wrote:

...
On Wed, May 21, 2025 at 12:35 PM Eugenio Perez Martin wrote:

...
On Wed, May 21, 2025 at 12:09 PM Stefano Brivio wrote:

...
On Tue, 20 May 2025 17:09:44 +0200 Eugenio Perez Martin wrote:

...
[...]

Now if I isolate the vhost kernel thread [1] I get way more performance as expected: - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 43.1 GBytes 37.1 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 43.1 GBytes 36.9 Gbits/sec receiver

After analyzing perf output, rep_movs_alternative is the most called function in the three iperf3 (~20%Self), passt.avx2 (~15%Self) and vhost (~15%Self)

Interesting... s/most called function/function using the most cycles/, I suppose.

Right!

...
So it looks somewhat similar to

https://archives.passt.top/passt-dev/20241017021027.2ac9ea53@elisabeth/

now?

Kind of. Below tcp_sendmsg_locked I don't see sk_page_frag_refill but skb_do_copy_data_nocache. Not sure if that means something, as it should not be affected by vhost.

...
...
But I don't see any of them consuming 100% of CPU in top: pasta consumes ~85% %CPU, both iperf3 client and server consumes 60%, and vhost consumes ~53%.

So... I have mixed feelings about this :). By "default" it seems to have less performance, but my test is maybe too synthetic.

Well, surely we can't ask Podman users to pin specific stuff to given CPU threads. :)

Yes but maybe the result changes under the right schedule? I'm isolating the CPUs entirely, which is not the usual case for pasta for sure :).

...
...
There is room for improvement with the mentioned optimizations so I'd continue applying them, continuing with UDP and TCP zerocopy, and developing zerocopy vhost rx.

That definitely makes sense to me.

Good!

...
...
With these numbers I think the series should not be merged at the moment. I could send it as RFC if you want but I've not applied the comments the first one received, POC style :).

I don't think it's really needed for you to spend time on semi-polishing something just to have an RFC if you're still working on it. I guess the implementation will change substantially anyway once you factor in further optimisations.

Agree! I'll keep iterating on this then.

Actually, if I remove all the taskset etc, and trust the kernel scheduler, vanilla pasta gives me: [pasta@virtlab716 ~]$ /home/passt/pasta --config-net iperf3 -c 10.6.68.254 -w 8M Connecting to host 10.6.68.254, port 5201 [ 5] local 10.6.68.20 port 40408 connected to 10.6.68.254 port 5201 [ ID] Interval Transfer Bitrate Retr Cwnd [ 5] 0.00-1.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 1.00-2.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 2.00-3.00 sec 3.12 GBytes 26.8 Gbits/sec 0 25.4 MBytes [ 5] 3.00-4.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 4.00-5.00 sec 3.10 GBytes 26.6 Gbits/sec 0 25.4 MBytes [ 5] 5.00-6.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 6.00-7.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 7.00-8.00 sec 3.09 GBytes 26.6 Gbits/sec 0 25.4 MBytes [ 5] 8.00-9.00 sec 3.08 GBytes 26.5 Gbits/sec 0 25.4 MBytes [ 5] 9.00-10.00 sec 3.10 GBytes 26.6 Gbits/sec 0 25.4 MBytes - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 31.0 GBytes 26.7 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 31.0 GBytes 26.5 Gbits/sec receiver

And with vhost-net : [pasta@virtlab716 ~]$ /home/passt/pasta --config-net iperf3 -c 10.6.68.254 -w 8M ... Connecting to host 10.6.68.254, port 5201 [ 5] local 10.6.68.20 port 46720 connected to 10.6.68.254 port 5201 [ ID] Interval Transfer Bitrate Retr Cwnd [ 5] 0.00-1.00 sec 4.17 GBytes 35.8 Gbits/sec 0 11.9 MBytes [ 5] 1.00-2.00 sec 4.17 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 2.00-3.00 sec 4.16 GBytes 35.7 Gbits/sec 0 11.9 MBytes [ 5] 3.00-4.00 sec 4.14 GBytes 35.6 Gbits/sec 0 11.9 MBytes [ 5] 4.00-5.00 sec 4.16 GBytes 35.7 Gbits/sec 0 11.9 MBytes [ 5] 5.00-6.00 sec 4.16 GBytes 35.8 Gbits/sec 0 11.9 MBytes [ 5] 6.00-7.00 sec 4.18 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 7.00-8.00 sec 4.19 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 8.00-9.00 sec 4.18 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 9.00-10.00 sec 4.18 GBytes 35.9 Gbits/sec 0 11.9 MBytes - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 41.7 GBytes 35.8 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 41.7 GBytes 35.7 Gbits/sec receiver

If I go the extra mile and disable notifications (it might be just noise, but...) [pasta@virtlab716 ~]$ /home/passt/pasta --config-net iperf3 -c 10.6.68.254 -w 8M ... Connecting to host 10.6.68.254, port 5201 [ 5] local 10.6.68.20 port 56590 connected to 10.6.68.254 port 5201 [ ID] Interval Transfer Bitrate Retr Cwnd [ 5] 0.00-1.00 sec 4.19 GBytes 36.0 Gbits/sec 0 12.4 MBytes [ 5] 1.00-2.00 sec 4.18 GBytes 35.9 Gbits/sec 0 12.4 MBytes [ 5] 2.00-3.00 sec 4.18 GBytes 35.9 Gbits/sec 0 12.4 MBytes [ 5] 3.00-4.00 sec 4.20 GBytes 36.1 Gbits/sec 0 12.4 MBytes [ 5] 4.00-5.00 sec 4.21 GBytes 36.2 Gbits/sec 0 12.4 MBytes [ 5] 5.00-6.00 sec 4.21 GBytes 36.1 Gbits/sec 0 12.4 MBytes [ 5] 6.00-7.00 sec 4.20 GBytes 36.1 Gbits/sec 0 12.4 MBytes [ 5] 7.00-8.00 sec 4.23 GBytes 36.4 Gbits/sec 0 12.4 MBytes [ 5] 8.00-9.00 sec 4.24 GBytes 36.4 Gbits/sec 0 12.4 MBytes [ 5] 9.00-10.00 sec 4.21 GBytes 36.2 Gbits/sec 0 12.4 MBytes - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 42.1 GBytes 36.1 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 42.1 GBytes 36.0 Gbits/sec receiver

So I guess the best is to actually run performance tests closer to real-world workload against the new version and see if it works better?

Well, that's certainly a possibility.

I'd say the biggest value for vhost-net usage in pasta is reaching throughput figures that are comparable with veth, with or without multithreading (keeping an eye on bytes per cycle, of course), with or without kernel changes, so that users won't need to choose between rootless and performance anymore.

It would also simplify things in Podman quite a lot (and to some extent in rootlesskit / Docker as well). We're pretty much there with virtual machines, just not quite with containers (which is somewhat ironic, but of course there's a good reason for that).

If we're clearly wasting cycles in vhost-net (because of the bounce buffer, plus something else perhaps?) *and* there's a somewhat possible solution for that in sight *and* the interface would change anyway, running throughput tests and polishing up the current version with a half-baked solution at the moment sounds a bit wasteful to me.

My point is that I'm testing a very synthetic scenario. If everybody agree this is close enough to real world ones, I'm ok to continue improving the edges we see. If not, maybe we're picking the wrong fruit even if it is low hand? Getting a table like [1] would give us light about this, especially if it is just a matter of running "make performance" or similar. Maybe we need to include longer queues? Focus on a given scenario? UDP goes better but TCP? Now more points about this scenario: 1) I don't see 100% CPU usage in any element: CPU% 84.2 passt.avx2 57.9 iperf3 57.2 iperf3 50.7 vhost-1805109 2) The most used (Self%) function in vhost is rep_movs_alternative, called from skb_copy_datagram_iter, so yes, ZeroCopy should help a lot here. Now, is "iperf3 -w 8M" representative? I'm sure ZC helps in this scenario, does it make it worse if we have small packets? Do we care? I'm totally ok with continuing trying with ZC, I just want to make sure we're not missing anything :). Thanks! [1] https://passt.top/passt/about/#performance_1 More details

...

But if one of those assumptions doesn't hold, or if you feel the need to consolidate the current status, perhaps polishing up the current version right now and actually evaluating throughput (as well as overhead) makes sense to me, yes.

-- Stefano

Stefano Brivio

10 Jun 10 Jun

5:29 p.m.

[Adding Paul as Podman developer] On Mon, 9 Jun 2025 11:59:21 +0200 Eugenio Perez Martin wrote:

...

On Fri, Jun 6, 2025 at 6:37 PM Stefano Brivio wrote:

...
On Fri, 6 Jun 2025 16:32:38 +0200 Eugenio Perez Martin wrote:

...
On Wed, May 21, 2025 at 12:35 PM Eugenio Perez Martin wrote:

...
On Wed, May 21, 2025 at 12:09 PM Stefano Brivio wrote:

...
On Tue, 20 May 2025 17:09:44 +0200 Eugenio Perez Martin wrote:

...
[...]

Now if I isolate the vhost kernel thread [1] I get way more performance as expected: - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 43.1 GBytes 37.1 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 43.1 GBytes 36.9 Gbits/sec receiver

After analyzing perf output, rep_movs_alternative is the most called function in the three iperf3 (~20%Self), passt.avx2 (~15%Self) and vhost (~15%Self)

Interesting... s/most called function/function using the most cycles/, I suppose.

Right!

...
So it looks somewhat similar to

https://archives.passt.top/passt-dev/20241017021027.2ac9ea53@elisabeth/

now?

Kind of. Below tcp_sendmsg_locked I don't see sk_page_frag_refill but skb_do_copy_data_nocache. Not sure if that means something, as it should not be affected by vhost.

...
...
But I don't see any of them consuming 100% of CPU in top: pasta consumes ~85% %CPU, both iperf3 client and server consumes 60%, and vhost consumes ~53%.

So... I have mixed feelings about this :). By "default" it seems to have less performance, but my test is maybe too synthetic.

Well, surely we can't ask Podman users to pin specific stuff to given CPU threads. :)

Yes but maybe the result changes under the right schedule? I'm isolating the CPUs entirely, which is not the usual case for pasta for sure :).

...
...
There is room for improvement with the mentioned optimizations so I'd continue applying them, continuing with UDP and TCP zerocopy, and developing zerocopy vhost rx.

That definitely makes sense to me.

Good!

...
...
With these numbers I think the series should not be merged at the moment. I could send it as RFC if you want but I've not applied the comments the first one received, POC style :).

I don't think it's really needed for you to spend time on semi-polishing something just to have an RFC if you're still working on it. I guess the implementation will change substantially anyway once you factor in further optimisations.

Agree! I'll keep iterating on this then.

Actually, if I remove all the taskset etc, and trust the kernel scheduler, vanilla pasta gives me: [pasta@virtlab716 ~]$ /home/passt/pasta --config-net iperf3 -c 10.6.68.254 -w 8M Connecting to host 10.6.68.254, port 5201 [ 5] local 10.6.68.20 port 40408 connected to 10.6.68.254 port 5201 [ ID] Interval Transfer Bitrate Retr Cwnd [ 5] 0.00-1.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 1.00-2.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 2.00-3.00 sec 3.12 GBytes 26.8 Gbits/sec 0 25.4 MBytes [ 5] 3.00-4.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 4.00-5.00 sec 3.10 GBytes 26.6 Gbits/sec 0 25.4 MBytes [ 5] 5.00-6.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 6.00-7.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 7.00-8.00 sec 3.09 GBytes 26.6 Gbits/sec 0 25.4 MBytes [ 5] 8.00-9.00 sec 3.08 GBytes 26.5 Gbits/sec 0 25.4 MBytes [ 5] 9.00-10.00 sec 3.10 GBytes 26.6 Gbits/sec 0 25.4 MBytes - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 31.0 GBytes 26.7 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 31.0 GBytes 26.5 Gbits/sec receiver

And with vhost-net : [pasta@virtlab716 ~]$ /home/passt/pasta --config-net iperf3 -c 10.6.68.254 -w 8M ... Connecting to host 10.6.68.254, port 5201 [ 5] local 10.6.68.20 port 46720 connected to 10.6.68.254 port 5201 [ ID] Interval Transfer Bitrate Retr Cwnd [ 5] 0.00-1.00 sec 4.17 GBytes 35.8 Gbits/sec 0 11.9 MBytes [ 5] 1.00-2.00 sec 4.17 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 2.00-3.00 sec 4.16 GBytes 35.7 Gbits/sec 0 11.9 MBytes [ 5] 3.00-4.00 sec 4.14 GBytes 35.6 Gbits/sec 0 11.9 MBytes [ 5] 4.00-5.00 sec 4.16 GBytes 35.7 Gbits/sec 0 11.9 MBytes [ 5] 5.00-6.00 sec 4.16 GBytes 35.8 Gbits/sec 0 11.9 MBytes [ 5] 6.00-7.00 sec 4.18 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 7.00-8.00 sec 4.19 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 8.00-9.00 sec 4.18 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 9.00-10.00 sec 4.18 GBytes 35.9 Gbits/sec 0 11.9 MBytes - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 41.7 GBytes 35.8 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 41.7 GBytes 35.7 Gbits/sec receiver

If I go the extra mile and disable notifications (it might be just noise, but...) [pasta@virtlab716 ~]$ /home/passt/pasta --config-net iperf3 -c 10.6.68.254 -w 8M ... Connecting to host 10.6.68.254, port 5201 [ 5] local 10.6.68.20 port 56590 connected to 10.6.68.254 port 5201 [ ID] Interval Transfer Bitrate Retr Cwnd [ 5] 0.00-1.00 sec 4.19 GBytes 36.0 Gbits/sec 0 12.4 MBytes [ 5] 1.00-2.00 sec 4.18 GBytes 35.9 Gbits/sec 0 12.4 MBytes [ 5] 2.00-3.00 sec 4.18 GBytes 35.9 Gbits/sec 0 12.4 MBytes [ 5] 3.00-4.00 sec 4.20 GBytes 36.1 Gbits/sec 0 12.4 MBytes [ 5] 4.00-5.00 sec 4.21 GBytes 36.2 Gbits/sec 0 12.4 MBytes [ 5] 5.00-6.00 sec 4.21 GBytes 36.1 Gbits/sec 0 12.4 MBytes [ 5] 6.00-7.00 sec 4.20 GBytes 36.1 Gbits/sec 0 12.4 MBytes [ 5] 7.00-8.00 sec 4.23 GBytes 36.4 Gbits/sec 0 12.4 MBytes [ 5] 8.00-9.00 sec 4.24 GBytes 36.4 Gbits/sec 0 12.4 MBytes [ 5] 9.00-10.00 sec 4.21 GBytes 36.2 Gbits/sec 0 12.4 MBytes - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 42.1 GBytes 36.1 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 42.1 GBytes 36.0 Gbits/sec receiver

So I guess the best is to actually run performance tests closer to real-world workload against the new version and see if it works better?

Well, that's certainly a possibility.

I'd say the biggest value for vhost-net usage in pasta is reaching throughput figures that are comparable with veth, with or without multithreading (keeping an eye on bytes per cycle, of course), with or without kernel changes, so that users won't need to choose between rootless and performance anymore.

It would also simplify things in Podman quite a lot (and to some extent in rootlesskit / Docker as well). We're pretty much there with virtual machines, just not quite with containers (which is somewhat ironic, but of course there's a good reason for that).

If we're clearly wasting cycles in vhost-net (because of the bounce buffer, plus something else perhaps?) *and* there's a somewhat possible solution for that in sight *and* the interface would change anyway, running throughput tests and polishing up the current version with a half-baked solution at the moment sounds a bit wasteful to me.

My point is that I'm testing a very synthetic scenario. If everybody agree this is close enough to real world ones, I'm ok to continue improving the edges we see. If not, maybe we're picking the wrong fruit even if it is low hand?

Getting a table like [1] would give us light about this, especially if it is just a matter of running "make performance" or similar. Maybe we need to include longer queues? Focus on a given scenario? UDP goes better but TCP?

Well, it's a matter of running ./run under tests (or 'make' there). Have you tried that with your patch? It's kind of representative in the sense that it uses several message sizes and different values for the sending window.

...

Now more points about this scenario: 1) I don't see 100% CPU usage in any element: CPU% 84.2 passt.avx2 57.9 iperf3 57.2 iperf3 50.7 vhost-1805109

Still, I bet we're using an awful amount of cycles compared to veth.

...

2) The most used (Self%) function in vhost is rep_movs_alternative, called from skb_copy_datagram_iter, so yes, ZeroCopy should help a lot here.

Now, is "iperf3 -w 8M" representative? I'm sure ZC helps in this scenario, does it make it worse if we have small packets? Do we care?

We don't care _a lot_ about small packets because we can typically use large packets, inbound and outbound, at least for TCP (bulk) transfers. But users are doing all sort of things with containers, including bulk transfers and VPN traffic over UDP, so we do, a bit. Again, the main value of using vhost-net, I think, is making "rootful" networking essentially unnecessary, or necessary just for niche use cases (say, non-TCP, non-UDP traffic, or macvlan-like cases). If there are relatively common use cases where pasta performs pretty bad compared to veth, we'll still need rootful networking. So, yes, it is representative, but not necessarily universal.

...

I'm totally ok with continuing trying with ZC, I just want to make sure we're not missing anything :).

In any case, it looks like vhost-net zero-copy is a bigger task than we thought, so, even if we don't reach a universal solution that makes rootful networking essentially unnecessary, but we have a big improvement ready, there's of course a lot of value in it. Your call... -- Stefano

Eugenio Perez Martin

11 Jun 11 Jun

9:04 a.m.

On Tue, Jun 10, 2025 at 5:29 PM Stefano Brivio wrote:

...

[Adding Paul as Podman developer]

On Mon, 9 Jun 2025 11:59:21 +0200 Eugenio Perez Martin wrote:

...
On Fri, Jun 6, 2025 at 6:37 PM Stefano Brivio wrote:

...
On Fri, 6 Jun 2025 16:32:38 +0200 Eugenio Perez Martin wrote:

...
On Wed, May 21, 2025 at 12:35 PM Eugenio Perez Martin wrote:

...
On Wed, May 21, 2025 at 12:09 PM Stefano Brivio wrote:

...
On Tue, 20 May 2025 17:09:44 +0200 Eugenio Perez Martin wrote:

> [...] > > Now if I isolate the vhost kernel thread [1] I get way more > performance as expected: > - - - - - - - - - - - - - - - - - - - - - - - - - > [ ID] Interval Transfer Bitrate Retr > [ 5] 0.00-10.00 sec 43.1 GBytes 37.1 Gbits/sec 0 sender > [ 5] 0.00-10.04 sec 43.1 GBytes 36.9 Gbits/sec receiver > > After analyzing perf output, rep_movs_alternative is the most called > function in the three iperf3 (~20%Self), passt.avx2 (~15%Self) and > vhost (~15%Self)

Interesting... s/most called function/function using the most cycles/, I suppose.

Right!

...
So it looks somewhat similar to

https://archives.passt.top/passt-dev/20241017021027.2ac9ea53@elisabeth/

now?

Kind of. Below tcp_sendmsg_locked I don't see sk_page_frag_refill but skb_do_copy_data_nocache. Not sure if that means something, as it should not be affected by vhost.

...
> But I don't see any of them consuming 100% of CPU in > top: pasta consumes ~85% %CPU, both iperf3 client and server consumes > 60%, and vhost consumes ~53%. > > So... I have mixed feelings about this :). By "default" it seems to > have less performance, but my test is maybe too synthetic.

Well, surely we can't ask Podman users to pin specific stuff to given CPU threads. :)

Yes but maybe the result changes under the right schedule? I'm isolating the CPUs entirely, which is not the usual case for pasta for sure :).

...
> There is room for improvement with the mentioned optimizations so I'd > continue applying them, continuing with UDP and TCP zerocopy, and > developing zerocopy vhost rx.

That definitely makes sense to me.

Good!

...
> With these numbers I think the series should not be > merged at the moment. I could send it as RFC if you want but I've not > applied the comments the first one received, POC style :).

I don't think it's really needed for you to spend time on semi-polishing something just to have an RFC if you're still working on it. I guess the implementation will change substantially anyway once you factor in further optimisations.

Agree! I'll keep iterating on this then.

Actually, if I remove all the taskset etc, and trust the kernel scheduler, vanilla pasta gives me: [pasta@virtlab716 ~]$ /home/passt/pasta --config-net iperf3 -c 10.6.68.254 -w 8M Connecting to host 10.6.68.254, port 5201 [ 5] local 10.6.68.20 port 40408 connected to 10.6.68.254 port 5201 [ ID] Interval Transfer Bitrate Retr Cwnd [ 5] 0.00-1.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 1.00-2.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 2.00-3.00 sec 3.12 GBytes 26.8 Gbits/sec 0 25.4 MBytes [ 5] 3.00-4.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 4.00-5.00 sec 3.10 GBytes 26.6 Gbits/sec 0 25.4 MBytes [ 5] 5.00-6.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 6.00-7.00 sec 3.11 GBytes 26.7 Gbits/sec 0 25.4 MBytes [ 5] 7.00-8.00 sec 3.09 GBytes 26.6 Gbits/sec 0 25.4 MBytes [ 5] 8.00-9.00 sec 3.08 GBytes 26.5 Gbits/sec 0 25.4 MBytes [ 5] 9.00-10.00 sec 3.10 GBytes 26.6 Gbits/sec 0 25.4 MBytes - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 31.0 GBytes 26.7 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 31.0 GBytes 26.5 Gbits/sec receiver

And with vhost-net : [pasta@virtlab716 ~]$ /home/passt/pasta --config-net iperf3 -c 10.6.68.254 -w 8M ... Connecting to host 10.6.68.254, port 5201 [ 5] local 10.6.68.20 port 46720 connected to 10.6.68.254 port 5201 [ ID] Interval Transfer Bitrate Retr Cwnd [ 5] 0.00-1.00 sec 4.17 GBytes 35.8 Gbits/sec 0 11.9 MBytes [ 5] 1.00-2.00 sec 4.17 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 2.00-3.00 sec 4.16 GBytes 35.7 Gbits/sec 0 11.9 MBytes [ 5] 3.00-4.00 sec 4.14 GBytes 35.6 Gbits/sec 0 11.9 MBytes [ 5] 4.00-5.00 sec 4.16 GBytes 35.7 Gbits/sec 0 11.9 MBytes [ 5] 5.00-6.00 sec 4.16 GBytes 35.8 Gbits/sec 0 11.9 MBytes [ 5] 6.00-7.00 sec 4.18 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 7.00-8.00 sec 4.19 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 8.00-9.00 sec 4.18 GBytes 35.9 Gbits/sec 0 11.9 MBytes [ 5] 9.00-10.00 sec 4.18 GBytes 35.9 Gbits/sec 0 11.9 MBytes - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 41.7 GBytes 35.8 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 41.7 GBytes 35.7 Gbits/sec receiver

If I go the extra mile and disable notifications (it might be just noise, but...) [pasta@virtlab716 ~]$ /home/passt/pasta --config-net iperf3 -c 10.6.68.254 -w 8M ... Connecting to host 10.6.68.254, port 5201 [ 5] local 10.6.68.20 port 56590 connected to 10.6.68.254 port 5201 [ ID] Interval Transfer Bitrate Retr Cwnd [ 5] 0.00-1.00 sec 4.19 GBytes 36.0 Gbits/sec 0 12.4 MBytes [ 5] 1.00-2.00 sec 4.18 GBytes 35.9 Gbits/sec 0 12.4 MBytes [ 5] 2.00-3.00 sec 4.18 GBytes 35.9 Gbits/sec 0 12.4 MBytes [ 5] 3.00-4.00 sec 4.20 GBytes 36.1 Gbits/sec 0 12.4 MBytes [ 5] 4.00-5.00 sec 4.21 GBytes 36.2 Gbits/sec 0 12.4 MBytes [ 5] 5.00-6.00 sec 4.21 GBytes 36.1 Gbits/sec 0 12.4 MBytes [ 5] 6.00-7.00 sec 4.20 GBytes 36.1 Gbits/sec 0 12.4 MBytes [ 5] 7.00-8.00 sec 4.23 GBytes 36.4 Gbits/sec 0 12.4 MBytes [ 5] 8.00-9.00 sec 4.24 GBytes 36.4 Gbits/sec 0 12.4 MBytes [ 5] 9.00-10.00 sec 4.21 GBytes 36.2 Gbits/sec 0 12.4 MBytes - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 42.1 GBytes 36.1 Gbits/sec 0 sender [ 5] 0.00-10.04 sec 42.1 GBytes 36.0 Gbits/sec receiver

So I guess the best is to actually run performance tests closer to real-world workload against the new version and see if it works better?

Well, that's certainly a possibility.

I'd say the biggest value for vhost-net usage in pasta is reaching throughput figures that are comparable with veth, with or without multithreading (keeping an eye on bytes per cycle, of course), with or without kernel changes, so that users won't need to choose between rootless and performance anymore.

It would also simplify things in Podman quite a lot (and to some extent in rootlesskit / Docker as well). We're pretty much there with virtual machines, just not quite with containers (which is somewhat ironic, but of course there's a good reason for that).

If we're clearly wasting cycles in vhost-net (because of the bounce buffer, plus something else perhaps?) *and* there's a somewhat possible solution for that in sight *and* the interface would change anyway, running throughput tests and polishing up the current version with a half-baked solution at the moment sounds a bit wasteful to me.

My point is that I'm testing a very synthetic scenario. If everybody agree this is close enough to real world ones, I'm ok to continue improving the edges we see. If not, maybe we're picking the wrong fruit even if it is low hand?

Getting a table like [1] would give us light about this, especially if it is just a matter of running "make performance" or similar. Maybe we need to include longer queues? Focus on a given scenario? UDP goes better but TCP?

Well, it's a matter of running ./run under tests (or 'make' there). Have you tried that with your patch? It's kind of representative in the sense that it uses several message sizes and different values for the sending window.

Yes but it freezes in my env. Copying the different windows, top-left: # tail -f --retry /home/passt/test/test_logs/context_unshare.log /home/passt/test/test_logs/context_ns.log tail: warning: --retry only effective for the initial open ==> /home/passt/test/test_logs/context_unshare.log <== unshare$ tail: cannot open '/home/passt/test/test_logs/context_ns.log' for reading: No such file or directory --- top-right: # while cat /tmp/passt-tests-7HpbEm/log_pipe; do :; done Test layout: single pasta instance with namespace. --- bottom-left: # tail -f --retry /home/passt/test/test_logs/context_host.log tail: warning: --retry only effective for the initial open host$ --- bottom-right: # tail -f --ontext_passt.logt/test/test_logs/c tail: warning: --retry only effective for the initial open passt$ --- And test/test_logs/test.log: === build/all

...

Build passt ? ! [ -e passt ] ? [ -f passt ] ...passed.

...

Build pasta ? ! [ -e pasta ] ? [ -h pasta ] ...passed.

...

Build qrap ? ! [ -e qrap ] ? [ -f qrap ] ...passed.

...

Build all ? ! [ -e passt ] ? ! [ -e pasta ] ? ! [ -e qrap ] ? [ -f passt ] ? [ -h pasta ] ? [ -f qrap ] ...passed.

...

Install ? [ -f /tmp/passt-tests-7HpbEm/build/all/prefix/bin/passt ] ? [ -h /tmp/passt-tests-7HpbEm/build/all/prefix/bin/pasta ] ? [ -f /tmp/passt-tests-7HpbEm/build/all/prefix/bin/qrap ] ? man -M /tmp/passt-tests-7HpbEm/build/all/prefix/share/man -W passt ? man -M /tmp/passt-tests-7HpbEm/build/all/prefix/share/man -W pasta ? man -M /tmp/passt-tests-7HpbEm/build/all/prefix/share/man -W qrap ...passed.

...

Uninstall ? ! [ -f /tmp/passt-tests-7HpbEm/build/all/prefix/bin/passt ] ? ! [ -h /tmp/passt-tests-7HpbEm/build/all/prefix/bin/pasta ] ? ! [ -f /tmp/passt-tests-7HpbEm/build/all/prefix/bin/qrap ] ? ! man -M /tmp/passt-tests-7HpbEm/build/all/prefix/share/man -W passt 2>/dev/null ? ! man -M /tmp/passt-tests-7HpbEm/build/all/prefix/share/man -W pasta 2>/dev/null ? ! man -M /tmp/passt-tests-7HpbEm/build/all/prefix/share/man -W qrap 2>/dev/null ...passed.

=== build/cppcheck ...skipped. === build/clang_tidy ...skipped. ---

...

...
Now more points about this scenario: 1) I don't see 100% CPU usage in any element: CPU% 84.2 passt.avx2 57.9 iperf3 57.2 iperf3 50.7 vhost-1805109

Still, I bet we're using an awful amount of cycles compared to veth.

...
2) The most used (Self%) function in vhost is rep_movs_alternative, called from skb_copy_datagram_iter, so yes, ZeroCopy should help a lot here.

Now, is "iperf3 -w 8M" representative? I'm sure ZC helps in this scenario, does it make it worse if we have small packets? Do we care?

We don't care _a lot_ about small packets because we can typically use large packets, inbound and outbound, at least for TCP (bulk) transfers. But users are doing all sort of things with containers, including bulk transfers and VPN traffic over UDP, so we do, a bit.

Again, the main value of using vhost-net, I think, is making "rootful" networking essentially unnecessary, or necessary just for niche use cases (say, non-TCP, non-UDP traffic, or macvlan-like cases). If there are relatively common use cases where pasta performs pretty bad compared to veth, we'll still need rootful networking.

So, yes, it is representative, but not necessarily universal.

...
I'm totally ok with continuing trying with ZC, I just want to make sure we're not missing anything :).

In any case, it looks like vhost-net zero-copy is a bigger task than we thought, so, even if we don't reach a universal solution that makes rootful networking essentially unnecessary, but we have a big improvement ready, there's of course a lot of value in it. Your call...

Got it, thanks!

Stefano Brivio

10:08 a.m.

On Wed, 11 Jun 2025 09:04:57 +0200 Eugenio Perez Martin wrote:

...

On Tue, Jun 10, 2025 at 5:29 PM Stefano Brivio wrote:

...
On Mon, 9 Jun 2025 11:59:21 +0200 Eugenio Perez Martin wrote:

...
Getting a table like [1] would give us light about this, especially if it is just a matter of running "make performance" or similar. Maybe we need to include longer queues? Focus on a given scenario? UDP goes better but TCP?

Well, it's a matter of running ./run under tests (or 'make' there). Have you tried that with your patch? It's kind of representative in the sense that it uses several message sizes and different values for the sending window.

Yes but it freezes in my env.

Copying the different windows, top-left: # tail -f --retry /home/passt/test/test_logs/context_unshare.log /home/passt/test/test_logs/context_ns.log tail: warning: --retry only effective for the initial open ==> /home/passt/test/test_logs/context_unshare.log <== unshare$ tail: cannot open '/home/passt/test/test_logs/context_ns.log' for reading: No such file or directory

[...]

=== build/cppcheck ...skipped.

=== build/clang_tidy ...skipped. ---

I guess you're missing some dependencies, nstool ('make assets', or 'make' under tests/), or similar. There's a test/README.md file listing most dependencies and also describing how to debug things, essentially: DEBUG=1 ./run and have a look at script.log. Patches for whatever issue you might be hitting in your setup are warmly welcome, of course. :) -- Stefano

Age (days ago)

Last active (days ago)

List overview

Download

10 comments

3 participants

participants (3)

Eugenio Perez Martin
Jason Wang
Stefano Brivio