Re: Question about Maglev algorithm
On Sat, Dec 29, 2018 at 06:03:51PM +0100, Aleksandar Lazic wrote: > I thought I have misunderstood the Idea behind maglev, thanks for > clarification. Found another mention of Maglev [Eis16] for high-level load balancing (between datacenters): https://landing.google.com/sre/sre-book/chapters/load-balancing-frontend/ -- Valentin
Re: Question about Maglev algorithm
Am 29.12.2018 um 07:41 schrieb Willy Tarreau: > On Fri, Dec 28, 2018 at 07:55:11PM +0100, Aleksandar Lazic wrote: >> Well as far as I understood the pdf one of the biggest difference is that >> Maglev is a distributed system where the consistent hash is for local system. > > No, not at all. The difference is that it's designed for packet processing > so they have to take care of connection tracking and per-packet processing > cost. From what I've read in the paper, it could be seen as a subset of > what we already do : > - server weights are not supported in Maglev (and very likely not needed) > - slow start is not supported > - server insertion/removal can be extremely expensive (O(N^2)) due to the > way they need to build the hash table for fast lookup > - no possibility for bounded load either > > It's really important to understand the different focus of the algorithm, > being packet-oriented instead of L7-oriented. This explains a number of > differences and choices. I think Maglev is excellent for what it does and > that our mechanism wouldn't be as fast if used on a per-packet basis. But > conversely, we already do the same and even much more by default because > we work at a different layer. I thought I have misunderstood the Idea behind maglev, thanks for clarification. > Willy Cheers Aleks
Re: Question about Maglev algorithm
On Fri, Dec 28, 2018 at 07:55:11PM +0100, Aleksandar Lazic wrote: > Well as far as I understood the pdf one of the biggest difference is that > Maglev is a distributed system where the consistent hash is for local system. No, not at all. The difference is that it's designed for packet processing so they have to take care of connection tracking and per-packet processing cost. From what I've read in the paper, it could be seen as a subset of what we already do : - server weights are not supported in Maglev (and very likely not needed) - slow start is not supported - server insertion/removal can be extremely expensive (O(N^2)) due to the way they need to build the hash table for fast lookup - no possibility for bounded load either It's really important to understand the different focus of the algorithm, being packet-oriented instead of L7-oriented. This explains a number of differences and choices. I think Maglev is excellent for what it does and that our mechanism wouldn't be as fast if used on a per-packet basis. But conversely, we already do the same and even much more by default because we work at a different layer. Willy
Re: Question about Maglev algorithm
Well as far as I understood the pdf one of the biggest difference is that Maglev is a distributed system where the consistent hash is for local system. What I think is if consistent hash uses the peers table for balancing it could be similar to Maglev, but I'm not a algo expert, just an Idea. I don't know if it's have any benefits for haproxy, I have seen this algo on envoy site and wanted to know what the experts here means about it :-) https://www.envoyproxy.io/docs/envoy/latest/intro/arch_overview/load_balancing/load_balancers Regards Aleks Ursprüngliche Nachricht Von: Aaron West Gesendet: 28. Dezember 2018 19:36:03 MEZ An: HAProxy Betreff: Re: Question about Maglev algorithm I've not used it yet with IPVS because I have nothing with a new enough Kernel (4.18+ I think), however, isn't this quite similar to HAProxy's consistent hash options? Aaron Loadbalancer.org
Re: Question about Maglev algorithm
I've not used it yet with IPVS because I have nothing with a new enough Kernel (4.18+ I think), however, isn't this quite similar to HAProxy's consistent hash options? Aaron Loadbalancer.org
Re: Question about Maglev algorithm
On Fri, Dec 28, 2018 at 07:11:24PM +0100, Aleksandar Lazic wrote: > Have anyone take a look into the Maglev algorithm ? > > This paper looks very interesting > https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/44824.pdf Seems to be in IPVS already: https://github.com/torvalds/linux/blob/master/net/netfilter/ipvs/ip_vs_mh.c -- Valentin
Question about Maglev algorithm
Hi. Have anyone take a look into the Maglev algorithm ? This paper looks very interesting https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/44824.pdf Regards Aleks