Thanks Miguel!
|------------>
| From: |
|------------>
>----------------------------------------------------------------------------------------------------------------------------------------|
|Miguel Delapaz/Endicott/i...@ibmus
|
>----------------------------------------------------------------------------------------------------------------------------------------|
|------------>
| To: |
|------------>
>----------------------------------------------------------------------------------------------------------------------------------------|
|IBMVM@LISTSERV.UARK.EDU
|
>----------------------------------------------------------------------------------------------------------------------------------------|
|------------>
| Date: |
|------------>
>----------------------------------------------------------------------------------------------------------------------------------------|
|12/11/2009 11:33 AM
|
>----------------------------------------------------------------------------------------------------------------------------------------|
|------------>
| Subject: |
|------------>
>----------------------------------------------------------------------------------------------------------------------------------------|
|Re: PATHMTU
|
>----------------------------------------------------------------------------------------------------------------------------------------|
|------------>
| Sent by: |
|------------>
>----------------------------------------------------------------------------------------------------------------------------------------|
|The IBM z/VM Operating System <IBMVM@LISTSERV.UARK.EDU>
|
>----------------------------------------------------------------------------------------------------------------------------------------|
Christy / Jonathan,
I've attempted to answer your questions/concerns below. If you have any
other questions, let me know.
> From what I can tell, the only place
> this comes in to play (generally) is for Hipersockets and similar
> private networks.
Any time there is a network with a smaller MTU than yours in between you
and your destination, Path MTU Discovery would come into play (assuming
you're sending large enough packets). So obviously networks with
significantly larger MTUs (like most HiperSockets networks) would it more
often.
> When I asked my network people about it, they
> said that the function has to be enabled and supported through all
> the hops in the network
This is not true. It improves the accuracy and speed of the Path MTU
algorithm if all the hops in the network send valid ICMP "fragmentation
needed" packets, but that can be done without enabling Path MTU Discovery
on the intermediate nodes. There are some operating systems that don't
send these packets (or send them without a suggested MTU value), but the VM
stack will adjust the MTU based on a table of well-known MTU values and
continue.
> that firewalls had to allow the ICMP
> traffic through.
Again, this makes the algorithm faster and more accurate, but it is not a
requirement. If we never receive the ICMP packet, the stack will
eventually adjust the MTU as above.
> There is also a rumor floating around that PATHMTU
> results in the “do not fragment’ bit always being set on. Which
> seems very scary and wrong.
This is not a rumor...but it's also not scary OR wrong. By forcing all
fragmentation to be done at the source host you are generally improving the
overall performance of the network because routers can spend their cycles
doing their primary job (routing traffic) rather than fragmenting packets
from a number of hosts. In fact, in IPv6 it is a requirement that ALL
fragmentation be done at the source node.
> Currently we have left it out so it
> is disabled and apparently
> defaults to Path MTU Discovery Aging Interval of 10 minutes
If Path MTU Discovery is disabled, the aging interval doesn't matter.
> We don't have a lot
> of IP traffic - is it OK to let it default
> to 10 minutes?
This is the value recommended by the RFC.
> Is there a way to determine what a 'good' value would
> be....?
In most cases, I don't think it matters too much. If you have a volatile
network, you would probably want the value lower in order to pick up MTU
increases faster. If you have a workload that sends small bursts of large
packets with a fair amount of time between the bursts, you may want the
value to be higher since there may be a (very) slight performance impact
while the algorithm determines the optimal MTU
> Is it that important...?
It definitely improves overall performance in some environments. If you're
in such an environment, then yes.
Regards,
Miguel Delapaz
z/VM Development
