On Jan 15, 2009, at 9:35 AM, Rakoczy, Dave wrote:
zLinux assigns the MTU size according to the IQD CHPID definition.
For sake of discussion lets say I set the CHPID to a Max Frame Size of
64K, that would give me an MTU size of 56K according to the Doc.
Where can I control the size of the packets I'll send across the
interface? In the Tape Blocksize / Record length as alluded to in the
Adam's previous note?
You can set the MTU manually with, say,
ifconfig hsi0 mtu 56000
If the guest is smart enough to set it to a reasonable default based
on the frame size that's probably what you want to do.
There are sysctl parameters you can tune to choose maximum packet read/
write sizes. See Mark Perry's note. If you are not memory
constrained, giving QDIO big buffers will also likely help speed
transfer speeds.
You can use the tracepath command to determine the maximum MTU size
across the whole path (which *should* be a direct-connection to your
zOS side, but probably worth verifying).
For your application, since you're using 256k tape blocks and 32k max
recls as your basic data chunk size, "as large as possible" is
probably the right answer for frame/MTU size...but by all means,
*measure performance* to see. We're also making the assumption that
the backup application just constructs network transactions in whole-
block or at least whole-record sizes. Without knowledge of the wire
protocol (which is likely not published) or tcpdumping what it's
doing, we don't really know that, and it is certainly not unheard of
for applications to try to be clever about network traffic on their
own (especially apps whose primary function is moving data across a
network quickly, like a backup client). So it may well end up being
the case that the app is performing a misguided optimization (for your
case--that is, FDR/Upstream may use behavior that is in fact the right
behavior for the majority of their customers but is wrong for your
situation) that nullifies the effects of moving the frame size up, and
you won't see a benefit.
We really don't know what relation the tape block and record size has
to the wire protocol. I, at least, am just guessing based on how I'd
design it: I've never looked that deeply into exactly HOW FDR/Upstream
moves its data around.
I feel like Bill Bitner: I must say, "it depends."
Adam
----------------------------------------------------------------------
For LINUX-390 subscribe / signoff / archive access instructions,
send email to lists...@vm.marist.edu with the message: INFO LINUX-390 or visit
http://www.marist.edu/htbin/wlvindex?LINUX-390
