Sorry about the vagueness there :)
Maybe the better way to put it is that each BMI _method_ (bmi_tcp,
bmi_mx, etc.) is expected to provide reliable, ordered delivery with
flow control. Neither the upper layer of BMI nor the users of BMI
provide those things, other than the fact that upper levels of PVFS will
at least timeout, cancel, and retry operations if they don't complete in
a reasonable amount of time. That's a pretty crude last resort for that
kind of scenario, though.
Most of the current BMI methods rely on their respective underlying
protocols for flow control, but if you were writing a method for
something like UDP that has no flow control at all, then you would have
to implement it yourself in the method.
The flow component (despite its name) doesn't really do flow control.
It maybe gets part of the way there simply by capping the number of
buffers (and therefore number of messages) in flight for each I/O
request as part of its pipelining mechanism, but it doesn't keep up with
tokens or anything that tell it about the state of the remote peer.
It's just trying to make reasonable use of local resources.
-Phil
On 01/07/2011 09:06 AM, Atchley, Scott wrote:
Phil,
bmi-design.tex has this paragraph:
"BMI provides reliability, message ordering, and flow control. If a particular
underlying protocol does not provide one of these features, then BMI is responsible for
implementing it."
Does BMI actually provide flow control? Does it rely on the method to provide
flow control? If so, I don't think I did anything with bmi_mx other than rely
on MX's flow control.
Or does Flow (or another upper layer protocol) provide flow control?
Thanks,
Scott
On Jan 5, 2011, at 10:59 AM, Atchley, Scott wrote:
Hi Phil,
Thanks, I'll take a look at it as well.
Scott
On Jan 5, 2011, at 10:52 AM, Phil Carns wrote:
Hi Scott,
I'm afraid so. The pvfs2 source tree also still has an old design
document in doc/design/bmi-design.tex.
-Phil
On 01/04/2011 03:09 PM, Atchley, Edward S. wrote:
Hi all,
Is this still the only BMI paper?
BMI: A Network Abstraction Layer for Parallel I/O
Scott
_______________________________________________
Pvfs2-users mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
_______________________________________________
Pvfs2-users mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
_______________________________________________
Pvfs2-users mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
_______________________________________________
Pvfs2-users mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
