| 1 |
On 03/03/2018 05:54 AM, Mick wrote: |
| 2 |
> UDP encapsulation as used for e.g. VPN does not suffer with the same |
| 3 |
> problem because it does not use the same transmission quality control |
| 4 |
> mechanism as TCP. |
| 5 |
|
| 6 |
I think it's fair to say that it doesn't suffer at the protocol (TCP / |
| 7 |
UDP) level. There is nothing to prevent higher application layer |
| 8 |
retransmissions from compounding things. |
| 9 |
|
| 10 |
> I am not sure if block device I/O protocols suffer the same problem - |
| 11 |
> I don't really know how the read/write SCSI commands are queued and |
| 12 |
> processed between host and guest OS. What I have noticed is abstraction |
| 13 |
> layers relating to partitioning schemes, e.g. good ol' primary Vs logical |
| 14 |
> partitions, make a difference *only* when the partition is initially |
| 15 |
> mounted, but not thereafter. |
| 16 |
|
| 17 |
I've always operated under the assumption that there was additional |
| 18 |
logic ~> complexity, thus it must be slightly slower. |
| 19 |
|
| 20 |
That being said, I've long held that the performance overhead is |
| 21 |
extremely likely negligible and can be ignored. At least unless you are |
| 22 |
trying to squeeze every bit of performance out of something. I.e. HPC |
| 23 |
or low power / low speed devices. |
| 24 |
|
| 25 |
|
| 26 |
|
| 27 |
-- |
| 28 |
Grant. . . . |
| 29 |
unix || die |
Archives