Hello, Thank you, I'll take a look at that. No academical research - I'm jjust curious :-) The cache interaction is of course big concern, but the general architecture described in the papers sidestep this by leaving the actual data copy for application core(s) - for both TX and RX. Please read the papers first. And of course this was meant for SMPs, I guess that for NUMA machines the interaction with memory subsystem might hamper scalability of such approach. For uniprocessors and smaller machines the traditional stack would most likely still be a better choice. Alex > ------------ Původní zpráva ------------ > Od: Matthew Faulkner <matthew.faulkner@xxxxxxxxx> > Předmět: Re: Dedicated core(s) for network stack? > Datum: 05.9.2010 23:25:35 > ---------------------------------------- > Ales, > > (Pimping my own work here) > > Not sure if you're doing this for academic work or not, either way > they might be interesting for you to read over. I basically concluded > excatly what Stephen just said at the end of his end. Dedicated > network processor is bad because of higher cache miss rates. The two > papers that might be of interest to you are: > > > Evaluating the Performance of Network Protocol Processing on Multi-core Systems > Architectural implications of performing network protocol processing > closer to the application (maybe slightly less interesting) > > If you're even more interested my PhD might be worth a read. It covers > the background on these topics in some detail. The can be found at > > http://matthew-faulkner.org/dump/thesis.pdf > > Hope this information is helpful > > Cheers. > > Dr. Matt Faulkner > > > > On 5 September 2010 21:22, Stephen Hemminger <shemminger@xxxxxxxxxx> wrote: > > On Sun, 05 Sep 2010 15:08:52 +0200 (CEST) > > ales-76@xxxxxxxxx wrote: > > > >> Hello, > >> > >> Recently I've read some nice articles: > >> > >> Shalev L. et al., IsoStack - Highly Efficient Network Processing on Dedicated > Cores > http://www.usenix.org/event/atc10/tech/full_papers/Shalev.pdf > >> Regnier G. et al., ETA: Experience with an Intel Xeon Processor as a Packet > Processing Engine > > http://www.hoti.org/archive/Hoti11_program/papers/hoti11_11_regnier_g.pdf > >> Regnier G. et al., TCP Onloading for Data Center Servers > > ftp://download.intel.com/technology/comms/perfnet/download/tcp_ieee_article.pdf > >> > >> Very interesting reading, I guess that many kernel people are familiar with > these papers. Anyway my question is: > >> > >> Any considerations (or even ongoing project) for implementing similar TCP/IP > stack architecture under Linux? > > > > No explicit project but it is an interesting idea. > > > >> It looks to me like very brigth idea, especially when number of cores on-chip > increases every year and the trend is clear. The benchmark seems to indicate > that isolating a core (or more) solely for network processing is a big win in > terms of performance. I am curious what is the opinion of kernel developers on > this. > > > > There is a myth that "network processing" is a some sort of monolithic entity > > with a single use case. For a typical server/client, there interactions of > > scheduler local processes and caching. If the processing is on a dedicated > > core, it guarantees a cache miss when the data is processed by the > > application. > > > > -- > > To unsubscribe from this list: send the line "unsubscribe linux-net" in > > the body of a message to majordomo@xxxxxxxxxxxxxxx > > More majordomo info at http://vger.kernel.org/majordomo-info.html > > > > > -- To unsubscribe from this list: send the line "unsubscribe linux-net" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
