Routing vs Switching and VLANs vs Flat

[Alex K]

Hi!

I would appreciate if you could help me answer if my concerns are valid.

I built a new network with 1 core + 7 access switches. All switches are 3750G. Core has an IP 10.100.1.1 and all access switches have their default route pointing to the core switch.

I decided to split a 10.100.0.0 network into /24 VLANs by logically organizing devices. 10.100.1.0/24 (Vlan 100) for all network devices, 10.100.2.0/24 (Vlan 2) for all servers, 10.100.13.0/24 (Vlan 13) for wireless APs, 10.100.16.0/23 (Vlan 16) for all DHCP devices, 10.100.7.0/24 (Vlan 7) for Voice etc. IP routing is enabled on all switches. Hosts in their subnets have gateways local to them, e.g. 10.100.16.25 has 10.100.16.1 as its gateway and 10.100.16.1 is the IP of Vlan 16. All ports belong to their corresponding vlans, except for the trunk ports that connect switches together.

I heard that generally switches do switching better than routing (although I also heard that only first packet to destination is being routed while all consecutive are being switched). My concern is if I ping a host 10.100.16.20 from a host on 10.100.2.0 network, all my ping replies vary between <1ms to 3-4ms. Does it mean that increased latency occurred because of routing between subnets? Should I be worried?

My other question is, if routing might increase the latency and potentially make network slower, should I simply make a flat /16 network and point all devices to my core 10.100.1.1 to eliminate routing? Basically, I'd get rid of all my vlans and simply organize hosts in logical subnets manually.

Network is not in production yet, so I still have some time to make final decision.

Thanks in advance for all your replies.

[devils_advocate]

Unless you have a larger number of Vlans/subnets (i.e. hundreds) then the 3750G is more than capable of routing between them.

The average response time is what you should be looking at, not the maximums.

In regards to your second question, no a large /16 network is not the way to go. The last thing you want is a huge broadcast domain like this.

[Alex K]

Thanks!

Is there anything else I should do to make the new network more efficient? Would it make sense to connect access switches to core using more then 1 cable to allocate more paths? How does that affect the overall throughput?

[pwwiddicombe]

This will sort of depend on the number of clients and their traffic patterns.  If you have a 1 gig uplink with all clients running 1 Gig, then it is quite possible you can periodically max out the uplink for short periods.  This may and may not be noticeable by the clients, depending on their applications.

It's easy enough to set up an etherchannel between core and client switches (and this CAN provide a level of fault tolerance if you run the cables in a different path); plus in future you could stack an additional core switch and have one uplink from EACH of them, to reduce the impact of failure.  Etherchannel will effectively double the effective bandwidth to the core; as half the clients will go down each link. 

Note, though, that you will still have the possibility of multiple heavy clients on the same link in the bundle; and their aggregate could still load one of the links in the bundle.