Question about MTU sizing

SlipperyPete · ‎12-07-2010

Over the years I've heard and read about MTU sizes and packet fragmentation, but as far as I can remember, in all the networks I've worked on, we've always left the configured MTU size as the default (1500), and applications seem to do just fine with that. Now I'm hearing about a customer that wants the MTU for their traffic to be set at 9000. Has anyone ran into applications that require a jumbo-sized frame like that? Also, if this customer wants their MTU set at 9000, wouldn't the MTU have to be set all along the path of whatever routers/switches their traffic crosses? Also, wouldn't the DF bit come into play? We manage a local site and some connections to remote locations, but we don't control whatever ISPs their traffic may have to traverse, and I'm thinking that even if we set our MTU at 9000, their packets will still get fragmented and/or dropped by other providers along the way. I'm just wondering what justification they would have for insisting on an MTU of 9000, and I suspect that it might be an older policy that they never bothered to re-evaluate.

Thanks.

Ganesh Hariharan · ‎12-07-2010

Over the years I've heard and read about MTU sizes and packet fragmentation, but as far as I can remember, in all the networks I've worked on, we've always left the configured MTU size as the default (1500), and applications seem to do just fine with that. Now I'm hearing about a customer that wants the MTU for their traffic to be set at 9000. Has anyone ran into applications that require a jumbo-sized frame like that? Also, if this customer wants their MTU set at 9000, wouldn't the MTU have to be set all along the path of whatever routers/switches their traffic crosses? Also, wouldn't the DF bit come into play? We manage a local site and some connections to remote locations, but we don't control whatever ISPs their traffic may have to traverse, and I'm thinking that even if we set our MTU at 9000, their packets will still get fragmented and/or dropped by other providers along the way. I'm just wondering what justification they would have for insisting on an MTU of 9000, and I suspect that it might be an older policy that they never bothered to re-evaluate.

Thanks.

Hi,

Jumbo frames are frames that are bigger than the standard Ethernet frame size, which is 1518 bytes (including Layer 2 (L2) header and FCS). The definition of frame size is vendor-dependent, as these are not part of the IEEE standard.Genarlly performace will be good if end server or client also support Jumbo frames if not then fragmentation will occur which results in degradation of performance.The performance problem arises when the MTU of one device is different than another.

Yes u need to enable jumbo frame support all over the path to have better performance however, there will be much ISPs offering Jumbo frame support.

Hope to Help !!

Ganesh.H

Remember to rate the helpful post

Scott Cannon · ‎12-07-2010

Jumbo frames have merit in some envrionments as they reduce the fragmenting and restructuing of packets and subsequently reduce latency and improve performance. In most instances this isnt an issue but when your talking hundreds of thousands of flows a second it can add up. You tend to see this more in SANs (iSCSI is a perfect example) but I'm yet to see it across a WAN. Ofcourse, that doesnt meant it doesnt happen

Rgds

Scott

darren.g · ‎12-07-2010

pweinhold wrote:

Over the years I've heard and read about MTU sizes and packet fragmentation, but as far as I can remember, in all the networks I've worked on, we've always left the configured MTU size as the default (1500), and applications seem to do just fine with that. Now I'm hearing about a customer that wants the MTU for their traffic to be set at 9000. Has anyone ran into applications that require a jumbo-sized frame like that? Also, if this customer wants their MTU set at 9000, wouldn't the MTU have to be set all along the path of whatever routers/switches their traffic crosses? Also, wouldn't the DF bit come into play? We manage a local site and some connections to remote locations, but we don't control whatever ISPs their traffic may have to traverse, and I'm thinking that even if we set our MTU at 9000, their packets will still get fragmented and/or dropped by other providers along the way. I'm just wondering what justification they would have for insisting on an MTU of 9000, and I suspect that it might be an older policy that they never bothered to re-evaluate.

Thanks.

The environment I work in has a number of applications which require jumbo frames - vendor specified, "This won't work without them" configurations. It's mainly in high bandwidth, high definition video environments that I've seen it, but yes, there are some applications which require (as opposed to "would like") jumbo frames.

And yes, you're correct that the jumbo frame size selected must be supported across the entire transmission path - if you're traversing WAN links, the WAN provider needs to have jumbo frame support enabled on your network - and if you're traversing Internet links you might as well forget it. Although it may be there are some VPN implementations which may "fool" the router interface into thinking a packet is unfragmented when it really has been fragmented as part of the VPN process - but I'm not going to offer up a vendor/implementation which will do this because I've never come across it, let alone tested it.

Are you sure they want this for external traffic? The usual demand for this is for some internal stuff - as i said, the commonest scenario I've seen it in is high definition video editing/recording environments (TV stations and the like), and it almost never has to leave the LAN and traverse the WAN.

Regards.

SlipperyPete · ‎12-09-2010

Yes, it's an application that will run across a WAN, but not the Internet. I think this WAN will be able to provide Jumbo frame support all the way across. This issue actually came up because these guys want to run their Jumbo-framed application across our connection to this WAN, and that connection happens to run through a 3750 switch. The 3750's only support up to a 9000 MTU, and they're doing some IP tunnelling that will add some overhead, so if their application sets the frame at around 9000, then their tunnel adds some overhead, there was a chance the 3750 would fragment it.

As it turned out, their engineers think that their app will work okay if they set the MTU at around 8900, which will ensure it hits our 3750 switch at less than 9000.

Still, it seems like poor design to use an application that requires large or Jumbo frames over a WAN. Even if the WAN is relatively stable, any lost packets and TCP will really bog down, with frames that big.

Thanks for everyone's input.

SlipperyPete · ‎12-09-2010

As a follow up to this discussion, does any one have any experience with Catalyst 3750 switches and MTU size limits?

I know you have to set the MTU as a global command (as opposed to per-interface) and it's capable of going up to 9000 bytes. However, one of our customers was concerned because they had read that a FE port on the 3750 wouldn't support Jumbo frame sizes. But we're using a the 12-port SFP model (WS-C3750G-12S-S), and every port on that model is a Gig port, so I think we'll be okay with setting its MTU at 9000.

Thanks.

darren.g · ‎12-09-2010

pweinhold wrote:

As a follow up to this discussion, does any one have any experience with Catalyst 3750 switches and MTU size limits?

I know you have to set the MTU as a global command (as opposed to per-interface) and it's capable of going up to 9000 bytes. However, one of our customers was concerned because they had read that a FE port on the 3750 wouldn't support Jumbo frame sizes. But we're using a the 12-port SFP model (WS-C3750G-12S-S), and every port on that model is a Gig port, so I think we'll be okay with setting its MTU at 9000.

Thanks.

The specification sheet located here

http://www.cisco.com/en/US/partner/prod/collateral/switches/ps5718/ps5023/product_data_sheet0900aecd80371991.html

indicates that Gigabit ethernet ports on the 3750 support a maximum ethernet frame size (jumbo frame size) of 9018 bytes (MTU 9000), but FE interfaces only support 1546 bytes.

So I reckon you'll be OK with the switches you have.

As far as your WAN links go - you really need to speak to the priovider - if it's a routed WAN then there's now ay you're going to get an MTU of 9000 across it - although 8900 should be fine - but if it's a layer 2-based WAN then you should be able to manage at least 8980.

Cheers