On Sat, Oct 07, 2023 at 10:56:31AM +0800, Wang Jinchao wrote: > On Wed, Oct 04, 2023 at 10:52:57AM -0400, Daniel Jordan wrote: > > Hi, > > > > On Wed, Sep 27, 2023 at 09:25:54PM +0800, Wang Jinchao wrote: > > > Hello, I have a few questions about the padata code I've been studying > > > recently: > > > > > > - Why does padata use the WQ_UNBOUND attribute of the workqueue? > > > > There's background in this series: > > https://lore.kernel.org/all/20190813005224.30779-1-daniel.m.jordan@xxxxxxxxxx/ > > > Summarizing, the use of the WQ_UNBOUND attribute is primarily based on performance > considerations. Is this understanding correct? Yes, it's entirely based on performance. In the tests I ran, padata did a lot better when it was free to run parallel works on any CPU in case some were busy. > > > Because I've noticed a significant maintenance cost related to CPUs. > > > Are there any specific benefits? > > > > Aside from what Steffen said about serialization, the pcrypt cpumasks > > can be set from sysfs to control where parallel and serial jobs run. > > > > > - In what scenarios is it necessary to specify a CPU for serial > > > execution, or is ensuring the order sufficient? > > > > I'm not sure that it's necessary. The way I read it, at least, it seems > > pcrypt uses cb_cpu to load balance serialization across all CPUs allowed > > in the serial cpumask. > > I recognize that the key issue isn't the capabilities it can offer, but rather > the underlying necessity for these capabilities. If the need isn't particularly > strong and the cost of implementing this capability is high, should we consider > omitting it? Well, yeah. In this case, I'm not sure it's necessary--it looks to me like it's not--but Steffen might have had a reason for cb_cpu that I'm not aware of. > As previously mentioned in another email, I'm seeking guidance on how to effectively > evaluate the timing of padata serialization in a network environment. > Could you provide some insights or pointers on this matter? It looks like Steffen has given some hints on testing with IPsec in another thread.
