[ofiwg] Potential agenda items for today

[Reese Faucette (rfaucett)]

> Last week I covered a proposal for handling 'credits'.  I don't think we've
> bottomed out on that.

Some discussion of "credits" would probably be useful, some different options discussed so far are:

a) have a well-known formula for computing credits used by an operation and let provider fill in constants (Sean's proposal unless I misrepresented)

b) have the provider keep its own formula for credit consumption and provide API to query "current credits" and "operation cost"

c) provide a "peek" style operation to check if the specified operation can complete or not

others?

My misgivings about (a) are 1) some new provider could come along requiring a modification to the "well-known" formula and have a hard time fitting in and 2) if this formula is used on the fly, it would generally entail a memory accesses with subsequent multiplication by zero, since provider constants that happen to be zero could not be optimized out.

Both (a) and (b) fall down when "ok to send" cannot be ascertained by a single scalar compare, such as a provider constrained by both buffer space and descriptor count.

So, I lean towards a b-style solution but concede there are providers where only (c) out of the above can work.  I'm ok with "if scalar compare inadequate, then credit system not supported, rely on EAGAIN"

-reese