The XGS 136 keeps the state of a UDP connection in hardware now, 2026-11-15
The XGS 136's keepalive asked whether a connection's state still existed and gave pf the coprocessor's counts, and
refreshed nothing: npuep_pf_touch, which restamps a state as a packet would, had no caller since the driver's
flow table began. So the state of a UDP connection in hardware lapsed at its own timeout - udp.multiple, 60
seconds - and the connection left hardware with it. Since the change the keepalive restamps a connection's states
when the coprocessor has counted traffic for it since the last restamp, at most every five seconds a connection
(counted in flow.ka_touched), and each restamp is told to pfsync (npuep_pf_sync), as the XGS 3300's driver does
(pfsync to a peer).
How it was measured
2026-11-15 from 06:32 to 08:36 UTC: the XGS 136 (vendor version 21) on the path of the lagg measurement - a jail on the FreeBSD test machine behind the second XGS 3300 (vendor version 22), through it to the XGS 136's Port7, and out of its Port11 to a second jail.
- The connection: UDP at 50 Mbit/s for 100 seconds from a fixed client port, routed by both appliances and in hardware on both.
- Read every 5 seconds: the XGS 136's state for it (
pfctl -ss -vv, its age) and whether it is in hardware there. A state replaced shows as its age falling back. - The builds: npuep.ko
520d842b(the module of phase 2's first part) before,03554cd2after, each loaded with reload.sh.
Results
| npuep.ko | the state's age, read every 5 s | in hardware on the XGS 136 |
|---|---|---|
520d842b, before |
0:58 at second 58, then 0:03 at second 64 - replaced | until second 58, then not again |
03554cd2, after |
0:04 to 1:46, rising at every read | the whole run |
- Before: in both runs the state was replaced 60 to 65 seconds in, and the stream went on through the XGS 136's host to its end.
- After: in three runs - one with
9a7510df, this change's first build, whose keepalive is the same - the state was never replaced, its age 1:44 to 1:47 at the last read, and the stream stayed in hardware throughout.


What is not measured
- pfsync with the XGS 136 carrying the connection: on this path the XGS 3300 carries it in hardware too, so two appliances would each hold a state for it and tell the other.
- TCP: a state's own timeout is a day, so it would lapse only after a day in hardware.
The discussion of this page is held in the project's GitHub Discussions. Showing it here loads giscus from giscus.app and GitHub, and writing in it needs a GitHub account - what that means.