lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
Open Source and information security mailing list archives
| ||
|
Message-ID: <F86E8680A7D9D74A99DC6345E35A6929501BA2CA@orsmsx503.amr.corp.intel.com> Date: Sat, 16 Jan 2010 02:53:15 -0800 From: "Waskiewicz Jr, Peter P" <peter.p.waskiewicz.jr@...el.com> To: David Miller <davem@...emloft.net> CC: "krkumar2@...ibm.com" <krkumar2@...ibm.com>, "netdev@...r.kernel.org" <netdev@...r.kernel.org>, "Kirsher, Jeffrey T" <jeffrey.t.kirsher@...el.com> Subject: RE: ixgbe: [RFC] [PATCH] Fix return of invalid txq >-----Original Message----- >From: David Miller [mailto:davem@...emloft.net] >Sent: Friday, January 15, 2010 1:06 AM >To: Waskiewicz Jr, Peter P >Cc: krkumar2@...ibm.com; netdev@...r.kernel.org; Kirsher, Jeffrey T >Subject: Re: ixgbe: [RFC] [PATCH] Fix return of invalid txq > >From: Peter P Waskiewicz Jr <peter.p.waskiewicz.jr@...el.com> >Date: Fri, 15 Jan 2010 01:00:20 -0800 > >> What I've been thinking of is more for the NUMA allocations per port. >> If we have, say 2 sockets, 8 cores a piece, then we have 16 CPUs. If >we >> assign a port to socket 0, I think the best use of resources is to >> allocate 8 Rx/Tx queues, one per core in that socket. If an >application >> comes from the other socket, we can have a table to map the other 8 >> cores from that socket into the 8 queues, instead of piling them all >> into one of the Tx queues. > >I fail to see how this can act substantially better than simply >feeding traffic evenly amongst whatever group of queues have >been configured. On systems with a large number of cores, and depending on how a NIC is allocated to a set of processors, there is a difference. If a port is allocated to socket three on a large NUMA system, then it could be CPUs 16-23 that it'd be using, and I'd have 8 Tx/Rx queues. I can either use the easy approach Krishna has, which could then execute two minus operations, or I can setup a lookup table in the driver on open(), and then I make a single indexed lookup. Using the lookup table makes the queue selection O(1) for whatever CPU/queue layout we come up with. Either way works though. I still think the table is the better way to go, because of the determinism for any system and NIC configuration/layout. The overhead of configuring the table is taken during open(), so it's not in the hotpath at all. Cheers, -PJ -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@...r.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Powered by blists - more mailing lists