On Wed, 8 Mar 2023 08:36:48 +0800 fengchengwen <fengcheng...@huawei.com> wrote:
> On 2023/3/8 0:06, Stephen Hemminger wrote: > > On Tue, 7 Mar 2023 17:35:32 +0800 > > fengchengwen <fengcheng...@huawei.com> wrote: > > > >> The syslog will add timestamp, but the syslog backend will re-write > >> timestamp, so > >> in the last, you can't find the real-timestamp of this log print. > >> sometimes it requires > >> to get real log time. > >> PS: we found it in our test environment because RR schedule hang too long > >> (similar question > >> also found: https://bugzilla.redhat.com/show_bug.cgi?id=1855447). > >> > >> So suggest add timestamp in syslog string also, and don't convert to > >> monotonic and just > >> print as normal format (just like syslog). > > > > > > Are you using systemd? > > Yes There is redhat bug about this: https://bugzilla.redhat.com/show_bug.cgi?id=991678 > > Never, never configure a DPDK application with real-time process priority. > > Polling model and RT don't mix. > > Maybe we should document them ? Part of previous discussion here: https://mails.dpdk.org/archives/dev/2021-April/203778.html In my experience running DPDK on isolated threads (cgroup or scheduler isolation) combined with remapping interrupts gives best response without lockup. I.e don't depend on scheduler to do the right thing. Instead ensure that each thread runs on dedicated CPU.
