network tweaking- Drop da Knowledge

RoundBoy

My current home network is :

2 tivos, and a wireless laptop connected to a 10/100 switch

1 server, and my main desktop, connected via 10/100/1000 switch. Both boxen have gig nics.

the two switches are also connected together.

While I am confident that everything is connected properly, I am not convinced I am using it at 100% potential. I enabled jumbo frames on both the server & desktop, but not via a yes/no, but as actually setting the MTU size.. i read that a 9000 jumbo frame size is really something like 9024 MTU with overhead, etc. Is this correct ?

What can i use to actually test the 'real' transfer speeds to various points on my network? I know I have 'netperf' for the linux server to tivos .. but my desktop i can only gauge how long it takes to transfer multiple gigs of video back and forth.

To sum up:

1) good tools to test network thoughput
2) MTU settings for 'jumbo frames' .. what should it actually be
3) Jumbo frames & a mixed 1000 / 100 network .. am i screwing female canines here ?

vonPoonBurGer

RoundBoy wrote: »

i read that a 9000 jumbo frame size is really something like 9024 MTU with overhead, etc. Is this correct ?

I think your MTU is the actual maximum packet size on the wire, including headers. So with an MTU of 9000, I think the maximum amount of data in a packet is 9000 minus the overhead, or 8976 bytes for standard TCP/IP. Whatever MTU you pick, just make sure it isn't actually getting fragmented somewhere along the way. If your switch is fragmenting the packet to smaller chunks, you're not getting the full benefit of jumbo frames. You can check for fragmentation between two jumbo hosts using this:

ping -t -l 8976 -f <IP of second jumbo-capable host>

If it says something about the packet needing fragmentation, but the Do Not Fragment flag is set, you know there's an issue. The exact syntax may depend on your OS, and what implementation of ping it has. I know the one above will work under Windows anyway.

You can also check to see if my theory about the MTU is correct. If 9000 bytes of data needs fragmentation but 8976 bytes doesn't, then I'm right. If 9000 bytes doesn't need fragmentation, then obviously I'm mistaken, and packet size on the wire is actually MTU plus overhead.

RoundBoy wrote: »

3) Jumbo frames & a mixed 1000 / 100 network .. am i screwing female canines here ?

I'm really not sure. It kind of depends on how you've implemented the coexistence of the two, and how well your router / switch supports that. From what I understand, there are basically two ways to mix jumbo and non-jumbo frames, on a port by port basis, or using two VLANs. Either way, when jumbo hosts talk to non-jumbo hosts, your switch needs to break the jumbo packets into 1500-byte chunks that your non-jumbo hosts can handle. Does your router / switch have some method to see its current load? Try copying some large files to one of the non-jumbo hosts, and check to see if your switch is getting pegged out in terms of processor or cache capacity. If so, you might want to reduce the jumbo frame size.

While it's not on your original list of questions, have you thought of doing any Quality of Service / traffic shaping on your network? While it's good to optimize your network at the media layers, it can also be useful to optimize the host layers that ride on top of the media layers. For example, you could prioritize latency-sensitive traffic (VOIP, gaming, etc.) over latency-insensitive bulk protocols (FTP, BitTorrent, etc.). I did that on my network a while back, and it's really nice not having to worry about whether my girlfriend remembered to cap the upload and download speeds on her torrents.

RoundBoy

excellent info. I am able to ping using the command above, but only up until a packet size of 8972 bytes. Was this 8976 a hard number, or an 'average' overhead?

I gleened this from linux settings somewhere for my ubuntu box, but a frame size of 9014 is an actual MTU of 9000 ... and since my gig nic has no direct settings for 'jumbo frames' on / off.. i set the frame size to be this 9014... is this correct ?

Pinging from the ubuntu box with a packet of 8976 shows (9004) .. which i assume is the actual packet size? If true, then 8972 seems fine from windows.

My network topology looks like :

cable modem --> router -->10/100 switch -->

the 10/100 splits to :
----> gig switch --> desktop, server
---> tivo a, tivo b, xbmc

It seems kike overkill network equipment wise, but I have ports in most every room, and the soon to be finished basement has a few multi-jack panels for computers going there.

My router is wireless (.11g) and is flashed to use better 3rd party firmware (including QOS). I am planning for the future, and my switches are unmananged . but the 10/100 switch has a large banwidth backplane.. and I have yet to see any issues with video streaming, and multi-gigabyte file transfers, and torrents running (all while playing wow

When I up the connection to FIOS, and start to dabble in VoIP, then I'll probably invest in a nice mananged switch.. but I truely think I am not stressing the network at all.

The tivo's have recently been patched with new drivers, but the usb 10/100 nic can only provide me with a transfer of 1-3 mb/sec .. and reports with custom kernels / gig nics / etc have people topping out at a MAX of 6 mb/sec .. hardly stressing a 10mbit port.

Since I can basically run packets outgoing from all networking devices in the house to a central place, is there a graphical tool , or whatever that lets me view traffic (ethereal ?)