Mark and I where playing with your test, and plotting the results I'm sharing the png's on a temp github here: https://github.com/rosenbaumalex/hashtest/ [I wasn't sure of a better place to share them] The README.md explains the port range we used, the 3 hash's used, and a line about the results. In general, the higher the 'noise' the worse the distribution is. It seems like Mark's hash suggestion (src*31 + dst) works best. then the folding one, and last the non-folding one. I am trying to cache a few switch related hash experts to get additional feedback. Alex On Fri, Feb 21, 2020 at 4:47 PM Tom Talpey <tom@xxxxxxxxxx> wrote: > > On 2/19/2020 8:04 PM, Mark Zhang wrote: > > On 2/20/2020 1:41 AM, Tom Talpey wrote: > >> On 2/19/2020 8:06 AM, Jason Gunthorpe wrote: > >>> On Wed, Feb 19, 2020 at 02:06:28AM +0000, Mark Zhang wrote: > >>>> The symmetry is important when calculate flow_label with DstQPn/SrcQPn > >>>> for non-RDMA CM Service ID (check the first mail), so that the server > >>>> and client will have same flow_label and udp_sport. But looks like it is > >>>> not important in this case. > >>> > >>> If the application needs a certain flow label it should not rely on > >>> auto-generation, IMHO. > >>> > >>> I expect most networks will not be reversible anyhow, even with the > >>> same flow label? > >> > >> These are network flow labels, not under application control. If they > >> are under application control, that's a security issue. > >> > > > > As Jason said application is able to control it in ipv6. Besides > > application is also able to control it for non-RDMA CM Service ID in ipv4. > > Ok, well I guess that's a separate issue, let's not rathole on > it here then. > > > Hi Jason, same flow label get same UDP source port, with same UDP source > > port (along with same sIP/dIP/sPort), are networks reversible? > > > >> But I agree, if the symmetric behavior is not needed, it should be > >> ignored and a better (more uniformly distributed) hash should be chosen. > >> > >> I definitely like the simplicity and perfect flatness of the newly > >> proposed (src * 31) + dst. But that "31" causes overflow into bit 21, > >> doesn't it? (31 * 0xffff == 0x1f0000) > > > > > I think overflow doesn't matter? We have overflow anyway if > > multiplicative is used. > > Hmm, it does seem to matter because dropping bits tilts the > distribution curve. Plugging ((src * 31) + dst) & 0xFFFFF into > my little test shows some odd behaviors. It starts out flat, > then the collisions start to rise around 49000, leveling out > at 65000 to a value roughly double the initial one (528 -> 1056). > It sits there until 525700, where it falls back to the start > value (528). I don't think this is optimal :-) > > Tom.
