I was wondering if anyone had looked at how, say, 2 (or more)
aggregated connections using 10Gb ethernet connections
compared as teamed vs. using them as bonded?
I know that a major hit in performance on both linux and windows
platforms is the time spent in interrupt processing (mostly in
deferred or SW interrupts, not HW interrupts).
With teamed connections on linux being in user space, I would
tend to think about how user-space file-system drivers perform
compared to drivers in the kernel. On another list, people were
talking about one of the limiting factors in NTFS performance on
linux was that it was a user-space driver, and as such, seemed
to generate noticeable performance limitations over native drivers.
Note, I use the example of 10G connections, as 1G connections
don't really stress today's processors on large reads/writes.
(Bad applications that use small, <64K buffers can... really
horrid are those that use 4K or smaller sizes at the application
level, the round-trip time is a killer, especially on local nets
w/Jumbo packets in the 9000+ byte range.
In case it isn't obvious, at those speeds, we are talking only
locally unencrypted links as encryption imposes extra latency
and bandwidth limitations.
Anyway, just wondering if teamed connections might go counter to
the the performance probs in file-systems?