Date: Sat, 19 Aug 2000 14:33:28 +1000 (EST) From: Stanley Hopcroft <Stanley.Hopcroft@IPAustralia.Gov.AU> To: isp@FreeBSD.ORG Subject: Throughput & Availability: Does anyone have experience with Trunking products (eg EtherChannel) ... ? Message-ID: <Pine.BSF.4.21.0008191402560.353-100000@stan>
next in thread | raw e-mail | index | archive | help
Dear Ladies and Gentlemen, I am writing about getting beter throughput and availability for servers (by having many NICs in each server), and ask would anyone please care to compare their experience with 1 Trunking products (ie VLANs between switch ports that connect to Servers) such as Foundry's trunking products and EtherChannel 2 Equal Cost Mult-path routing (with or without a routing core). Writers to this list have commented favourably on the equal cost multi-path option, but that option seems to be notably absent from switch vendor literature; they only talk trunks (perhaps so they can sell more switch ports. See for example the Foundry ISP Co-Location and Co-Hosting "case study") The theoretical pluses and minuses of each seem to me to be Factor Trunk Multi-Path n x Throughput No (<= 2 x ) ~ = (n = number of NICs) eg 4 100 TX NICs => 200 Mbps => 400 Mbps Auto Failover Yes Yes (by switch) (by routing process) Switch Ports == NICs No (always 2 x) Yes dual or quad trunk any number of ports 2 or 4 ports Layer 2 (802.1q) Yes No Layer 3 No Yes Standard if 802.1q not ISL No ==> Available for FreeBSD No Yes Available for famous brand servers Yes (Sun, NT, AIX) Yes (maybe with Gated or if you've got the RRAS) right OS, driver etc All switches No (must have correct Yes firmware etc) Same NICs Yes No in server (supported with trunking drivers etc) TCP reordering No Yes Needs an L3 Core No Yes Needs a routing process No Yes Per server (or many defaults) My conclusion is that trunking products provide something like a statistical load balancing (half traffic uses one NIC, the remainder the other) and does not increase the client - server link capacity. It is therefore a feeble method of increasing server throughput even though it does improve link availality ? The only practical disadvantage of equal cost multi path are running routing processes on servers, having the TCP driver re order packets, and needing an L3 switch core (so that client traffic is fowarded to the server by all of the routes through the server NICs). Would any one like to comment on this, or better still let me know why trunking is a better proposal ? Alternatvely, why would anyone use Trunking (EtherChannel) ? I suppose Multi-link PPP is completely out of the question because no switch supports it ? Does anyone use the Foundry ServerIron only to provide better server throughput ? (the "Virtual IP" corresponds to many NICs in one server) Thank you, Yours sincerely. S Hopcroft Network Specialist IP Australia +61 2 6283 3189 +61 2 6281 1353 FAX To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-isp" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?Pine.BSF.4.21.0008191402560.353-100000>