08-28-2007 11:55 AM - edited 03-05-2019 06:09 PM
Aim: create a server farm block for about 400 servers that have to dual home into the 2 switches that would comprise of this block
I already have a pair of spare 6500's which I can reuse for a server farm block. It currently has the SUP2 and the WS-X6548-GE-TX modules which I believe have an oversubscription rate of 8:1
Option1:
I can upgrade the Sup2 to Sup720 on these 6513's and use this as my server farm block. I'm aware that the 6548 wouldn't truly leverage the 720 fabric but I don't think the next option does any better
Option2:
- Trade in all of the above equipment and buy new 4500 switches to serve as the server farm block. However, I checked out the line cards for the 4500 and the best GE module seems to be the WS-X4548-GB-RJ45 which also has a 8:1 oversubscription
Is one option better than the other? Obviously, with option 1, I have to replace fewer components hence, it would be an easier sell on my side
Thank you
08-28-2007 12:23 PM
Hi
With a Sup720 you have an inbuit Switch Fabric Module and as the WS-X6548-GE-TX is fabric enabled at least you won't be on the 32Gbps shared bus.
The WS-X6548-GE-TX is not a server farm module. As you say it is oversubscribed 8:1.
Assuming you don't have a separate SFM already in your chassis then going with the Sup720 will at least move you forward but without upgrading your modules you are not really benefitting in terms of throughput.
Are all servers equal ie. do some need a lot more throughput than others ?
The 4500 is definitely not a better option.
HTH
Jon
08-28-2007 12:47 PM
1) There's a WS-X6500-SFM2 in the switch. Will upgrading to the SUP720 provide some benefit at all or none if I stick with these 6548 blades?
2) Servers are unequal and some are using higher bandwidth than others
3) Can you think of any other switching options for servers? There's the 4948 which may have a better oversub rate but I really need a chassis option
4) Do the 6748's have an oversub of 1:1? I know they connect to the fabric using 40gbps which is less than the 48 ports can use up, i.e. 48gbps
Thanks so much
08-28-2007 01:21 PM
1) sup720 provides fabric, so SFM redundant.
3) How about stacks of 3750-Es? (For raw performance, only a full up 6500, i.e. with sup720 and DFC, has better numbers.)
2) Place higher bandwidth usage servers on 3750-E that has uplink - then Stackwise bandwidth not used
4) Yes the 6748 is oversubscribed. (Actually remember seeing a tech note, so is the 4 port 10 gig card, very slightly.)
08-28-2007 10:46 PM
Hi
1) In terms of pure throughput if you have an SFM then very little really other than the fact you can free up a slot.
However Sup720 is future proofing
2 & 3) Then it may make more sense to look at purchasing modules that are not as badly oversubscribed for the high throughput servers but be aware that you cannot run the 67xx modules without a Sup720
4) The 6748's have 2 x 20Gb connections to the switch fabric. As you say they can support 48 x 1Gb connections so you are oversubscribed bit nothing like the 6548-GE-TX.
HTH
Jon
08-29-2007 11:28 AM
My understanding is that you will run into additional problems when using a 6513 chassis even with a sup720 because not all cards will get full 40Gb bandwidth. The 6509 is currently the largest chassis you can get full bandwidth using the 67xx series cards (and of course the sup720).
Paul
08-29-2007 04:20 PM
2) If you're using a card that has port groups, i.e. some group of ports map to a bottleneck, you use less ports to reduce the oversubscription even down to 1:1. E.g. only use 19 gig ports of the first (1..24) half and 19 gig ports of the second (25..48) half of the 6748's 48 ports. Then your only using 38 gig.
If you know the static bandwidth usage per server, you can manually balance server connected ports to the port groups to obtain what ever static load you want on the port group. E.g. if you have 6 gig ports on the card that feed to 1 gig, you could connect one server that normally uses 50%, one that uses 30%, and two that use 10%, to minimize the probability of oversubscription of the bottleneck.
Of course, doing all these port allocations is quite a bit of work and if the server bandwidth usage changes, you need to recompute. Usually not worth doing except for critically important servers.
Find answers to your questions by entering keywords or phrases in the Search bar above. New here? Use these resources to familiarize yourself with the community: