From optical-networks@mailman.isi.edu Mon Mar 4 19:45:22 2002 From: optical-networks@mailman.isi.edu (Joe Touch) Date: Mon, 04 Mar 2002 11:45:22 -0800 Subject: [optical-networks] FWIW, two interesting refs Message-ID: <3C83CED2.30700@isi.edu> "High-Speed Networking: A Systematic Approach to High-Bandwidth Low-Latency Communication," J. Sterbenz, J. Touch, John Wiley & Sons, April 2001, ISBN: 0471330361. (FWIW, Alan has a copy of this) This explains how protocols and high-speed interact. http://www.isi.edu/touch/oi97/ This is a meeting I co-chaired in 1997 at Bert Hui's request, to investigate where Optical Internetting was going. I hope they're useful. Joe From optical-networks@mailman.isi.edu Wed Mar 13 03:11:32 2002 From: optical-networks@mailman.isi.edu (Michelle C. Hauer) Date: Tue, 12 Mar 2002 19:11:32 -0800 Subject: [optical-networks] Questions for OFC Talk Message-ID: <3C8EC364.9070602@tsotech.com> I am preparing my slides to present at the Optical Fiber Communications conference next week for the optical correlators paper and I have some questions for the ISI folks: 1. I know you've told me this before, but why do we only care about 24 bits of the destination IP address? I recall something about the first 8 bits always being the same for historical reasons (because they represent tier 1 addresses or something). 2. How many MIPS or GIPS do modern routers operate at (for fast routing lookups). You guys mentioned in your presentation that it takes ~100 instructions to search for an entry. Can I use the MIPS number to get an estimate for how long this lookup takes? (I know that depends on how long an instruction really takes, which can be more than 1 clock cycle, but perhaps it is okay for estimation purposes). I would like to be able to say, with some confidence, that lookups presently take microseconds while our optical bypass can reduce this to nanoseconds (only limited by the optical switching time, not the correlators - which are ultimately limited by the size of the wavelength of the light). 3. How often do we anticipate that the configuration of the correlators will have to be re-computed? (Given that the bits the correlators are configured to match must be updated as the routing table evolves). I'm not necessarily going to put these details onto the slides, but I'd like to know the answers in case I get related questions. Thanks! If you'd rather just answer these questions over the phone: Pager: 310-781-1962 (also accepts short emails: mp@tsotech.com) Home phone (after 7:45pm tonight): 310-322-2755 Work phone (tomorrow): 310-647-8680 Michelle From optical-networks@mailman.isi.edu Wed Mar 13 04:55:21 2002 From: optical-networks@mailman.isi.edu (Joe Touch) Date: Tue, 12 Mar 2002 20:55:21 -0800 Subject: [optical-networks] Questions for OFC Talk References: <3C8EC364.9070602@tsotech.com> Message-ID: <3C8EDBB9.30508@isi.edu> Michelle C. Hauer wrote: > I am preparing my slides to present at the Optical Fiber Communications > conference next week for the optical correlators paper and I have some > questions for the ISI folks: > > 1. I know you've told me this before, but why do we only care about 24 > bits of the destination IP address? I recall something about the first > 8 bits always being the same for historical reasons (because they > represent tier 1 addresses or something). Actually, it's the other reason - the LAST 8 bits are always used to address the host within a network. Only the network part is routed. In the old days (pre early 90's, something called CIDR, or Class-less Inter-Domain Routing), there were 3 groups: Class A: starts with 0, the next 7 bits used to route to the network, and the remaining 24 bits used to address hosts within the network Class B: starts with 10, the next 14 bits used to route to the network, and the remaining 16 bits used to address hosts within the network Class C: starts with 100, the next 21 bits used to route to the network, remaining 8 bits used to address hosts within the network CIDR eliminated these distinctions, but, regardless, only the first 24 bits are used to route in the core. > 2. How many MIPS or GIPS do modern routers operate at (for fast routing > lookups). You guys mentioned in your presentation that it takes ~100 > instructions to search for an entry. Can I use the MIPS number to get > an estimate for how long this lookup takes? Not quite. Routers don't have CPUs - they use dedicated hardware on each input line, so MIPS doesn't make sense here. Our '100 instruction' comparison is more for host-based processors (i.e., workstation-based). > (I know that depends on how > long an instruction really takes, which can be more than 1 clock cycle, > but perhaps it is okay for estimation purposes). I would like to be > able to say, with some confidence, that lookups presently take > microseconds while our optical bypass can reduce this to nanoseconds > (only limited by the optical switching time, not the correlators - which > are ultimately limited by the size of the wavelength of the light). The right numbers here are that current routers can handle 100,000's of packets per second (PPS). An old number was 200,000 for a fast router, it's probably closer to 6-800,000 PPS. Which means 1 microsecond per lookup - they don't have parallel lookups and pipelining, so the time per lookup is just the inverse of the rate. > 3. How often do we anticipate that the configuration of the correlators > will have to be re-computed? (Given that the bits the correlators are > configured to match must be updated as the routing table evolves). Routing tables get updated on the order of minutes (3-5 minutes). Packet flows last around that long as well. If we reload every 5 minutes, that'd be fine. Note that the reload time only has to cost us about 5% downtime to be essentially nonexistent. Let me know if that helps... Joe From optical-networks@mailman.isi.edu Wed Mar 13 07:21:27 2002 From: optical-networks@mailman.isi.edu (Joseph Bannister) Date: Tue, 12 Mar 2002 23:21:27 -0800 Subject: [optical-networks] Questions for OFC Talk In-Reply-To: <3C8EC364.9070602@tsotech.com> Message-ID: <000101c1ca9a$045c59a0$0100007f@pow> -----Original Message----- From: optical-networks-admin@ISI.EDU [mailto:optical-networks-admin@ISI.EDU]On Behalf Of Michelle C. Hauer Sent: Tuesday, March 12, 2002 7:12 PM To: Optical Networks Mailing List Subject: [optical-networks] Questions for OFC Talk I am preparing my slides to present at the Optical Fiber Communications conference next week for the optical correlators paper and I have some questions for the ISI folks: 1. I know you've told me this before, but why do we only care about 24 bits of the destination IP address? I recall something about the first 8 bits always being the same for historical reasons (because they represent tier 1 addresses or something). *** The 24 bits of significance come from the fact that IP address formats are 32 bits in the pattern net_num.host_num where net_num may be 8, 16, or 24 bits, depending on whether they are the so-called class A, B, or C addresses. I'm attaching a PowerPoint slide of the formats. 2. How many MIPS or GIPS do modern routers operate at (for fast routing lookups). You guys mentioned in your presentation that it takes ~100 instructions to search for an entry. Can I use the MIPS number to get an estimate for how long this lookup takes? (I know that depends on how long an instruction really takes, which can be more than 1 clock cycle, but perhaps it is okay for estimation purposes). I would like to be able to say, with some confidence, that lookups presently take microseconds while our optical bypass can reduce this to nanoseconds (only limited by the optical switching time, not the correlators - which are ultimately limited by the size of the wavelength of the light). *** High-end routers these days do not implement address lookup in a general-purpose processor. Instead, lookup is largely realized by some ASIC or combination of ASICs. I'd say that the upper end of routing is around 50,000,000 packets per second, which translates to a lookup time of less than 20 seconds per lookup (in a table of over 150,000 address-prefix entries). Remember to mention that pressure for faster lookup comes from two fronts: rapidly increasing routing-table size AND rapidly increasing link speed; so the the time window to look up a packet's address keeps dropping dramatically. 3. How often do we anticipate that the configuration of the correlators will have to be re-computed? (Given that the bits the correlators are configured to match must be updated as the routing table evolves). ***This is for Joe Touch to answer. I'm not necessarily going to put these details onto the slides, but I'd like to know the answers in case I get related questions. Thanks! If you'd rather just answer these questions over the phone: Pager: 310-781-1962 (also accepts short emails: mp@tsotech.com) Home phone (after 7:45pm tonight): 310-322-2755 Work phone (tomorrow): 310-647-8680 Michelle *** If you'd like to include us in a dry run, please let us know when it would be convenient to do so. _______________________________________________ Optical-networks mailing list Optical-networks@mailman.isi.edu http://mailman.isi.edu/mailman/listinfo/optical-networks From optical-networks@mailman.isi.edu Wed Mar 13 15:36:53 2002 From: optical-networks@mailman.isi.edu (Michelle C. Hauer) Date: Wed, 13 Mar 2002 07:36:53 -0800 Subject: [optical-networks] Questions for OFC Talk Message-ID: <200203131536.g2DFapR16346@gamma.isi.edu> Thank you both for your responses. They are very helpful. Michelle > -----Original Message----- > From: optical-networks-admin@ISI.EDU > [mailto:optical-networks-admin@ISI.EDU]On Behalf Of Michelle C. Hauer > Sent: Tuesday, March 12, 2002 7:12 PM > To: Optical Networks Mailing List > Subject: [optical-networks] Questions for OFC Talk > I am preparing my slides to present at the Optical Fiber Communications > conference next week for the optical correlators paper and I have some > questions for the ISI folks: > 1. I know you've told me this before, but why do we only care about 24 > bits of the destination IP address? I recall something about the first > 8 bits always being the same for historical reasons (because they > represent tier 1 addresses or something). > *** The 24 bits of significance come from the fact that IP address formats > are 32 bits in the pattern net_num.host_num where net_num may be 8, 16, or > 24 bits, depending on whether they are the so-called class A, B, or C > addresses. I'm attaching a PowerPoint slide of the formats. > 2. How many MIPS or GIPS do modern routers operate at (for fast routing > lookups). You guys mentioned in your presentation that it takes ~100 > instructions to search for an entry. Can I use the MIPS number to get > an estimate for how long this lookup takes? (I know that depends on how > long an instruction really takes, which can be more than 1 clock cycle, > but perhaps it is okay for estimation purposes). I would like to be > able to say, with some confidence, that lookups presently take > microseconds while our optical bypass can reduce this to nanoseconds > (only limited by the optical switching time, not the correlators - which > are ultimately limited by the size of the wavelength of the light). > *** High-end routers these days do not implement address lookup in a > general-purpose processor. Instead, lookup is largely realized by some > ASIC > or combination of ASICs. I'd say that the upper end of routing is around > 50,000,000 packets per second, which translates to a lookup time of less > than 20 seconds per lookup (in a table of over 150,000 address-prefix > entries). Remember to mention that pressure for faster lookup comes from > two fronts: rapidly increasing routing-table size AND rapidly increasing > link speed; so the the time window to look up a packet's address keeps > dropping dramatically. > 3. How often do we anticipate that the configuration of the correlators > will have to be re-computed? (Given that the bits the correlators are > configured to match must be updated as the routing table evolves). > ***This is for Joe Touch to answer. > I'm not necessarily going to put these details onto the slides, but I'd > like to know the answers in case I get related questions. > Thanks! > If you'd rather just answer these questions over the phone: > Pager: 310-781-1962 (also accepts short emails: mp@tsotech.com) > Home phone (after 7:45pm tonight): 310-322-2755 > Work phone (tomorrow): 310-647-8680 > Michelle > *** If you'd like to include us in a dry run, please let us know when it > would be convenient to do so. > _______________________________________________ > Optical-networks mailing list > Optical-networks@mailman.isi.edu > http://mailman.isi.edu/mailman/listinfo/optical-networks > _______________________________________________ > Optical-networks mailing list > Optical-networks@mailman.isi.edu > http://mailman.isi.edu/mailman/listinfo/optical-networks ********************************************************* Michelle C. Hauer Research Assistant Optical Communications Laboratory Dept. of Electrical Engineering University of Southern California Tel: 213/740.1488 Fax: 213/740.8729 Email: hauer@usc.edu http://oclab.usc.edu/ *********************************************************