Re: [AusNOG] Deep buffer switches and CDN networks
Hi Tom, Appreciate both of your comments. Perfectly timed video from Netflix too, thanks for pointing me towards it! @Paul Wilkins I think that's what confused me the most around deeper buffers. I get how they could help with burst absorption for applications which sometimes might contend with each other. But for an openconnect box (or other type of CDN) I can only assume North-South traffic is constantly going to be coming in thick and fast. No amount of buffering is going to fix what is essentially a big to little pipe scenario. The bigger buffers may come in useful for clocking (PPM) differences between like interfaces, but this would from what I understand only really help to make the benchmarks/marketing material look better. Thanks all! Regards, Jason. On Fri, 24 May 2019 at 23:31, Tom Paseka wrote: > Timely presentation from Netflix On this exact topic. > https://ripe78.ripe.net/archives/video/128/ > > On Thu, May 23, 2019 at 7:52 AM Tom Paseka wrote: > >> everything depends on your application and how you're moving traffic. >> >> if you lots of east-west flow (between equal speed interfaces, especially >> in many to one) you'll need buffers. If you're doing north to south traffic >> with interface change, you'll likely need buffers. >> >> The choice here might not have been for deep buffer, but for other >> capabilities (forwarding, route table size, etc). Dave mentions from a >> question there is no east-west traffic, everything is south-north. >> >> On Wed, May 22, 2019 at 5:46 AM Jason Leschnik wrote: >> >>> Hi Noggers, >>> >>> I just finished watching the NANOG presentation of Netflix >>> openconnect[1], I noticed that their core switch of choice was an Arista >>> 7500E which is a deep buffer switch. I remember seeing a lot of comments >>> around buffer bloat for deep buffer switches. Would this be considered an >>> acceptable use case for this type of switch in the DC? >>> >>> Has anyone got experience with ideal switch types (shallow, deep >>> buffers) for edge CDN network deployments? >>> >>> [1] - https://www.youtube.com/watch?v=tbqcsHg-Q_o >>> >>> Regards, >>> Jason. >>> ___ >>> AusNOG mailing list >>> AusNOG@lists.ausnog.net >>> http://lists.ausnog.net/mailman/listinfo/ausnog >>> >> ___ AusNOG mailing list AusNOG@lists.ausnog.net http://lists.ausnog.net/mailman/listinfo/ausnog
Re: [AusNOG] Deep buffer switches and CDN networks
Deep buffers of transit routers do not add bandwidth, nor do they improve your bandwidth/delay product. The only use case I've ever found for deep buffers is remote sites on low bandwidth links, where with the right QoS maps you can shunt bulk transfer traffic aside, meaning user traffic eg voice, isn't contending with say email. Once you have 10 or more concurrent users, the law of large numbers means deep buffers just take longer to fill before buffer drop. Buffer drop never means you need deeper buffers, it means you need more bandwidth. Perhaps there's a case for shunting update traffic between CDN caches, but again, this relies on there being a benefit in delaying CDN cache update packets from a user traffic peak to a user traffic trough, (where the delay between peak and trough < 1s), the absence of user traffic freeing bandwidth for the cache update. Again, the law of large numbers makes this unlikely. Kind regards Paul Wilkins On Fri, 24 May 2019 at 23:32, Tom Paseka wrote: > Timely presentation from Netflix On this exact topic. > https://ripe78.ripe.net/archives/video/128/ > > On Thu, May 23, 2019 at 7:52 AM Tom Paseka wrote: > >> everything depends on your application and how you're moving traffic. >> >> if you lots of east-west flow (between equal speed interfaces, especially >> in many to one) you'll need buffers. If you're doing north to south traffic >> with interface change, you'll likely need buffers. >> >> The choice here might not have been for deep buffer, but for other >> capabilities (forwarding, route table size, etc). Dave mentions from a >> question there is no east-west traffic, everything is south-north. >> >> On Wed, May 22, 2019 at 5:46 AM Jason Leschnik wrote: >> >>> Hi Noggers, >>> >>> I just finished watching the NANOG presentation of Netflix >>> openconnect[1], I noticed that their core switch of choice was an Arista >>> 7500E which is a deep buffer switch. I remember seeing a lot of comments >>> around buffer bloat for deep buffer switches. Would this be considered an >>> acceptable use case for this type of switch in the DC? >>> >>> Has anyone got experience with ideal switch types (shallow, deep >>> buffers) for edge CDN network deployments? >>> >>> [1] - https://www.youtube.com/watch?v=tbqcsHg-Q_o >>> >>> Regards, >>> Jason. >>> ___ >>> AusNOG mailing list >>> AusNOG@lists.ausnog.net >>> http://lists.ausnog.net/mailman/listinfo/ausnog >>> >> ___ > AusNOG mailing list > AusNOG@lists.ausnog.net > http://lists.ausnog.net/mailman/listinfo/ausnog > ___ AusNOG mailing list AusNOG@lists.ausnog.net http://lists.ausnog.net/mailman/listinfo/ausnog
Re: [AusNOG] Deep buffer switches and CDN networks
Timely presentation from Netflix On this exact topic. https://ripe78.ripe.net/archives/video/128/ On Thu, May 23, 2019 at 7:52 AM Tom Paseka wrote: > everything depends on your application and how you're moving traffic. > > if you lots of east-west flow (between equal speed interfaces, especially > in many to one) you'll need buffers. If you're doing north to south traffic > with interface change, you'll likely need buffers. > > The choice here might not have been for deep buffer, but for other > capabilities (forwarding, route table size, etc). Dave mentions from a > question there is no east-west traffic, everything is south-north. > > On Wed, May 22, 2019 at 5:46 AM Jason Leschnik wrote: > >> Hi Noggers, >> >> I just finished watching the NANOG presentation of Netflix >> openconnect[1], I noticed that their core switch of choice was an Arista >> 7500E which is a deep buffer switch. I remember seeing a lot of comments >> around buffer bloat for deep buffer switches. Would this be considered an >> acceptable use case for this type of switch in the DC? >> >> Has anyone got experience with ideal switch types (shallow, deep buffers) >> for edge CDN network deployments? >> >> [1] - https://www.youtube.com/watch?v=tbqcsHg-Q_o >> >> Regards, >> Jason. >> ___ >> AusNOG mailing list >> AusNOG@lists.ausnog.net >> http://lists.ausnog.net/mailman/listinfo/ausnog >> > ___ AusNOG mailing list AusNOG@lists.ausnog.net http://lists.ausnog.net/mailman/listinfo/ausnog
Re: [AusNOG] Deep buffer switches and CDN networks
everything depends on your application and how you're moving traffic. if you lots of east-west flow (between equal speed interfaces, especially in many to one) you'll need buffers. If you're doing north to south traffic with interface change, you'll likely need buffers. The choice here might not have been for deep buffer, but for other capabilities (forwarding, route table size, etc). Dave mentions from a question there is no east-west traffic, everything is south-north. On Wed, May 22, 2019 at 5:46 AM Jason Leschnik wrote: > Hi Noggers, > > I just finished watching the NANOG presentation of Netflix openconnect[1], > I noticed that their core switch of choice was an Arista 7500E which is a > deep buffer switch. I remember seeing a lot of comments around buffer bloat > for deep buffer switches. Would this be considered an acceptable use case > for this type of switch in the DC? > > Has anyone got experience with ideal switch types (shallow, deep buffers) > for edge CDN network deployments? > > [1] - https://www.youtube.com/watch?v=tbqcsHg-Q_o > > Regards, > Jason. > ___ > AusNOG mailing list > AusNOG@lists.ausnog.net > http://lists.ausnog.net/mailman/listinfo/ausnog > ___ AusNOG mailing list AusNOG@lists.ausnog.net http://lists.ausnog.net/mailman/listinfo/ausnog
[AusNOG] Deep buffer switches and CDN networks
Hi Noggers, I just finished watching the NANOG presentation of Netflix openconnect[1], I noticed that their core switch of choice was an Arista 7500E which is a deep buffer switch. I remember seeing a lot of comments around buffer bloat for deep buffer switches. Would this be considered an acceptable use case for this type of switch in the DC? Has anyone got experience with ideal switch types (shallow, deep buffers) for edge CDN network deployments? [1] - https://www.youtube.com/watch?v=tbqcsHg-Q_o Regards, Jason. ___ AusNOG mailing list AusNOG@lists.ausnog.net http://lists.ausnog.net/mailman/listinfo/ausnog