Re: [ofa-general] [PATCH] ipoib/cm: use common CQ for all TX QPs
Michael S. Tsirkin wrote: >> Quoting Pradeep Satyanarayana <[EMAIL PROTECTED]>: >> Subject: Re: [ofa-general] [PATCH] ipoib/cm: use common CQ for all TX QPs >> >> Michael S. Tsirkin wrote: >>> Use common CQ for all TX QPs: keep a per-device counter of outstanding TX >>> WRs, >>> and stop the interface when this counter reaches the send queue size, to >>> avoid >>> CQ overruns. This should help reduce the number of interrupts for >>> bi-directional traffic (such as TCP). >> CQ overrun generates asynchronous errors. If CQ overruns are avoided, how >> does that >> reduce interrupts? >> So, the number of interrupts are reduced by switching to polling mode on the TX side too; not by preventing CQ overruns. >> On a side note, as a further extension to this idea, would it be worth >> considering >> sendq_size and recvq_size being device specific in the future, instead of >> being >> module specific? > > How would you do this without resize QP support in low leve drivers? Yes, that would have to be a prerequisite. > >>> This helps fix "driver is hogging interrupts" errors reported for IPoIB >>> send side. >>> See e.g. https://bugs.openfabrics.org/show_bug.cgi?id=508 >>> >> In bug 508, timer ticks were lost because some device was hogging >> interrupts. >> What is the association between dropping packets and hogging interrupts? >> Can you expand on that? > > This patch fixes the hogging interrupts problem by polling TX completions > from NAPI thread. > Right, this likely solves the hogging interrupts problem. But, wasn't dmesg flooded with dropping packets messages. That issue might still not be resolved. Pradeep ___ general mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
Re: [ofa-general] [PATCH] ipoib/cm: use common CQ for all TX QPs
> Quoting Pradeep Satyanarayana <[EMAIL PROTECTED]>: > Subject: Re: [ofa-general] [PATCH] ipoib/cm: use common CQ for all TX QPs > > Michael S. Tsirkin wrote: > > Use common CQ for all TX QPs: keep a per-device counter of outstanding TX > > WRs, > > and stop the interface when this counter reaches the send queue size, to > > avoid > > CQ overruns. This should help reduce the number of interrupts for > > bi-directional traffic (such as TCP). > > CQ overrun generates asynchronous errors. If CQ overruns are avoided, how > does that > reduce interrupts? > > On a side note, as a further extension to this idea, would it be worth > considering > sendq_size and recvq_size being device specific in the future, instead of > being > module specific? How would you do this without resize QP support in low leve drivers? > > This helps fix "driver is hogging interrupts" errors reported for IPoIB > > send side. > > See e.g. https://bugs.openfabrics.org/show_bug.cgi?id=508 > > > In bug 508, timer ticks were lost because some device was hogging interrupts. > What is the association between dropping packets and hogging interrupts? > Can you expand on that? This patch fixes the hogging interrupts problem by polling TX completions from NAPI thread. -- MST ___ general mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
Re: [ofa-general] [PATCH] ipoib/cm: use common CQ for all TX QPs
Michael S. Tsirkin wrote: > Use common CQ for all TX QPs: keep a per-device counter of outstanding TX WRs, > and stop the interface when this counter reaches the send queue size, to avoid > CQ overruns. This should help reduce the number of interrupts for > bi-directional traffic (such as TCP). CQ overrun generates asynchronous errors. If CQ overruns are avoided, how does that reduce interrupts? On a side note, as a further extension to this idea, would it be worth considering sendq_size and recvq_size being device specific in the future, instead of being module specific? > > This helps fix "driver is hogging interrupts" errors reported for IPoIB send > side. > See e.g. https://bugs.openfabrics.org/show_bug.cgi?id=508 > In bug 508, timer ticks were lost because some device was hogging interrupts. What is the association between dropping packets and hogging interrupts? Can you expand on that? Pradeep ___ general mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
