[infinispan-dev] Google load balancer design

[Radim Vansa]

On 03/22/2016 08:24 AM, Emmanuel Bernard wrote:
> http://blog.acolyer.org/2016/03/21/maglev-a-fast-and-reliable-software-network-load-balancer/
>
> Interesting read. Some data points I extracted.
>
> 350 ns per parquet. Ensuring a 300us during load.

Regrettably JGroups does not bypass Linux kernel (yet). It sounds like 
the forwarder is written for the particular NIC model (or maybe reuses 
the driver internally), though it's in userspace.

>
> Some math I was not aware of:
>
> [quote]
>
> --
>
> Let /N/be the size of a VIP’s backend pool. A table is constructed 
> with/M/entries, where/M/is a prime number and/M > 100 * N/, which will 
> ensure at most a 1% difference in hash space assigned to backends.
>
> --
>
> Might be interesting in our virtual nodes logic. The segment size 
> might not be too friendly here though.

No virtual nodes since Infinispan 5.2.0.

R.

>
> They do use try to ensure relative stability of the targeted backend 
> for a given key even when the pool of backend changes (e.g. one server 
> down)
>
> Emmanuel
>
>
>
> _______________________________________________
> infinispan-dev mailing list
> infinispan-dev at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/infinispan-dev


-- 
Radim Vansa <rvansa at redhat.com>
JBoss Performance Team