From mboxrd@z Thu Jan 1 00:00:00 1970 From: Yevgeny Kliteynik Subject: Re: [PATCH 3/3 v2] opensm: connect switches in tree - implemented in up/down Date: Sun, 06 Dec 2009 10:35:45 +0200 Message-ID: <4B1B6CE1.4050802@dev.mellanox.co.il> References: <4AF160D8.9040801@dev.mellanox.co.il> <20091126154643.GA13817@me> Reply-To: kliteyn-LDSdmyG8hGV8YrgS2mwiifqBs+8SCbDb@public.gmane.org Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: <20091126154643.GA13817@me> Sender: linux-rdma-owner-u79uwXL29TY76Z2rM5mHXA@public.gmane.org To: Sasha Khapyorsky Cc: Linux RDMA List-Id: linux-rdma@vger.kernel.org Hi Sasha, Sasha Khapyorsky wrote: > On 13:09 Wed 04 Nov , Yevgeny Kliteynik wrote: >> This patch implements connect_switches option in up/down >> routing. Also, connect_roots is now handled as a special >> case of connect_switches. >> >> The idea is the following: when clearing hops, preserve >> the entries for switches that are above the highest leaf >> in the tree. >> So if the highest leaf in the tree has rank N, preserve >> hops to all the switches with ranks 0 to (N-1). >> When connecting roots (--connect_roots option), just set >> N to 1. > > Would this affect multicast routing in sense of a credit loop > generation? Since I sent these patches, I had it running in various simulations and setups, and there are couple of fundamental issue with this approach. Basically, what I did here (in up/dn) is wrong. Not only for multicast, but for unicast too. It interferes with the usual up/down paths routing. I will issue V3 of the patches, and it will be only connect_roots implementation for fat-tree with the two small remarks that you found. As for the general algorithm for connecting switches, the whole approach of the connect_roots option and of what I was lately trying to do with connect_switches option is wrong. There is no need to connect all the roots to each other. There's also no need to connect all the switches to each other. What we need is a connection between all the *managed* switches (extended port 0) to all the other switches in the fabric in both directions. This option should *replace* the connect_roots option functionality (though we can leave connect_roots for backward compatibility). I'll work on the algorithm, but clearly it won't be ready for OFED 1.5. If you have any thoughts about this idea, I'd be happy to hear them. -- Yevgeny -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org More majordomo info at http://vger.kernel.org/majordomo-info.html