Den mån 25 maj 2020 kl 10:03 skrev Marc Roos <M.Roos@xxxxxxxxxxxxxxxxx>: > > I am interested. I am always setting mtu to 9000. To be honest I cannot > imagine there is no optimization since you have less interrupt requests, > and you are able x times as much data. Every time there something > written about optimizing the first thing mention is changing to the mtu > 9000. Because it is quick and easy win. > > This sort of assumes you are not using interrupt coalescing network cards, because if you do, you can get something like hundreds of packets in one single IRQ*, already checksummed and stripped and in recent cards (10-25-40GE) even delivered into the cpu L3 cache by the time you get the int, so if they were 1500 or 9000 on the wire doesn't matter much by then. Even in the bad old days of software handling of all parts packet-related, many things (like mbuf allocations) were optimized for 1500, so 9k packets became just a multiple of a number of 1500 bytes chunks taken from a pool of network buffers anyhow. I'm not trying to shoot down the 9k-vs-1500 idea, but doing a benchmark will give you lots more facts than airing things that are easy to imagine but really doesn't have a huge impact because hw manufacturers worked around things like this a long time ago. If your tests say you win x%, then use it by all means. I'm just not thinking that 10/25/40G networks are so filled that the frame overheads really matter as a matter of % of the packet sizes and the cards offload most of the work to strip the overhead out, so the computer won't notice it was ever there. *) SysKonnect cards had this around 2003, just to get a feeling for what "modern ethernet cards" means in this context. -- May the most significant bit of your life be positive. _______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx
 |