Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
On Thu, 2009-12-10 at 16:52 +0100, Lars Ellenberg wrote: > This likely has been correct at some point in time, or for 100 MBit/s. > > In my experience, bonding two 1GBit links can yield about 1.7 to 1.8, > adding the third drops back to about 1.0 but adds latency, > and adding a forth will deteriorate to less then a single link. > > This is for a _sinlge_ tcp bulk transfer. Of course, the sum total over > all transfers on those bonded links will be higher. > > But for the single tcp bulk transfer, in my experience, more than two > 1GBit links in bonding-rr are counterproductive. > > May be highly variable with different NICs, boards, > processors and bus and processor speeds. I've tested the network and there is no evident problems with netcat transfers. Also, i don't see how high disk IO on secondary can be result of the network interfaces, because according to atop, interfaces are loaded only approx 1-3% each. I've tried LACP too, and I have exact the same problems with drbd. When DRBD is disconnected, device gives me double to triple performance, also changing from protocol C to A or B gives no significant improvement. -- | Jakov Sosic | ICQ: 28410271 | PGP: 0x965CAE2D | ================================================================= | |