Re: MaxRecvPDULength question

[Bill Studenmund <wrstuden@wasabisystems.com>]

On Tue, 21 May 2002, Paul Koning wrote:

> Excerpt of message (sent 21 May 2002) by Luben Tuikov:
> > The key name ``MaxRecvPDULength'' doesn't imply
> > <MaxRecvDataLength>? Why is it considered as
> > such then?
> >
> > I suggest that ``MaxRecvPDULength'' be set to mean
> > the maximum PDU length, NOT the maximum _data_ length.
> >
> > Reason: It is _faster_ to get the PDU off the the TCP
> > data stack in _one_ (system) call, rather than two (system) calls.
> > This _would_ improve implementaions, more so for user-space such.

Unfortunately I don't think we can do that. The problem is that we still
would need to know how much data were sent. Mainly as we don't HAVE to
send the maximum amount of data. :-)

> > But lets still keep the size of the PDU a power of 2.
> > (This would imply that the non-existent MaxRecvDataLength
> > to MaxRecvPDULength - 48.)
> >
> > Comments?
>
> I can't quite figure out what you're after, but it doesn't sound
> good.
>
> There are two parameters of interest: the size of the entire iSCSI
> PDU, and the size of the data portion of the PDU.
>
> It doesn't matter a whole lot which of the two you specify, because
> the difference is simply the iSCSI header size.  I find it useful to
> think in terms of the data length, because that relates to alignment
> of data going into the storage machinery, etc.

I agree.

> Are you proposing to rename the key so its name is
> "MaxRecvDataLength"?  Yes, that would be more accurate, but why bother
> at this point?
>
> The overall PDU size current is NOT a power of two given the current
> defaults, and should not be, because that would make the data length a
> weird value.
>
> As for your comment about 1 vs. 2 system calls, I don't understand
> what system calls you're talking about.  The performance argument is
> not meaningful because it may only be done once per system boot, or at
> the very worst once per login.

Luben was wanting to make it so the WHOLE PDU (header and data both) could
be read with one system call.

Take care,

Bill