switch performance WS-C3560G-48TS-S

[peterve]

Hi,

does anyone have an idea what the effective throughput would be in the following scenario

Server generating data in memory, copying it to another server (into memory)

Both machines are connected to the same Cisco switch (WS-C3560G-48TS-S) using a single cat6 ethernet cable between server and switch.

What would be the impact of adding interfaces to the server and use multiple cables between the servers and the switch

What would be the impact if I decide to use TOE host bus adapters instead of regular network interface cards

What would the results look like if was doing the test from SCSI harddrive to SCSI harddrive (in other words : will the harddrives, both 15000 rpm) be the bottleneck or not ?

thanks

PS : the machines sit next to each other; there's nothing running on the server except for the OS. Cable distance : 2meter

--------------------------------------------------------------------
How can I believe in God when just last week I got my tongue caught in the roller of an electric typewriter?
---------------------------------------------------------------------

 

[lambent]

use multiple cables between the servers and the switch"

You may try to use "teaming" in this case and set FastEther Channel on the switch. In this case the multiple physical connections form a single logical connection to the switch, providing redundancy on NIC and cable level.

 

[peterve]

I know, but my question is more related to the performance of the solution. What will be the effective throughput I'm getting ?
what is the impact of setting nics in teaming mode, or use TOE HBA's ?

--------------------------------------------------------------------
How can I believe in God when just last week I got my tongue caught in the roller of an electric typewriter?
---------------------------------------------------------------------

 

[lambent]

In your case if you use teaming, you'll have an aggregated bandwidth of 400Mbps full-duplex with load-sharing on the 2 physical links.

 

[peterve]

what is the effective throughput in that case ?

--------------------------------------------------------------------
How can I believe in God when just last week I got my tongue caught in the roller of an electric typewriter?
---------------------------------------------------------------------

 

[lambent]

Since there's load-sharing between the 2 physical links and each links support up to 100Mbps inbound and 100Mbps outbound traffic rate, you'll have an aggregated 400Mbps maximum bandwidth. Excluding all the TCP/IP overheads, effectively you may have a maximum of 60% - 70% of the aggregated bandwidth, that's about 240Mbps to 280Mbps aggregated bandwidth. Just imagine sending a 30MB to 35MB file in 1 second.

 

[peterve]

ok - that's what I had in mind... but I have to be sure (based upon facts, not assumptions)

Has anyone else tested this ? Anyone able to provide me real test results ?

--------------------------------------------------------------------
How can I believe in God when just last week I got my tongue caught in the roller of an electric typewriter?
---------------------------------------------------------------------

 

[ADB100]

If you attempted copying files between 2 machines, each with EtherChannel configured the maximum throughput you would theoretically see would be 100Mbps (200Mbps full duplex). This is due to the load-sharing algorythm that runs on both the switch and the host. There are different algorythms that can be enabled to choose which port in the channel is used to transmit a packet - these can be based on MAC address, IP Address or Layer-4 port number (with IP packets only). The Host's load-sharing algorythm is usually either MAC or IP based only, I haven't seen any adapters that use Layer-4 information but there may be some?

Between 2 hosts there isn't any variation in the variables used to determine the port in the channel to use so all traffic will traverse the same physical link. If you have multiple data transfers concurrently to many different hosts the algorythm is more useful as it will load-balance each traffic-flow based on the unique variables each one will present.

That being said the host itself is usually the limiting factor since it is unlikley to be able to keep up with the throughput requests. Your best bet is to attempt your data transfer measurement test using a single port between 2 hosts using an Ethernet cross-over cable, this way you will have a baseline. Following that enable a channel between the 2 hosts (2 cross-over cables etc) and test again - you shouldn't see any improvement and the interface statistics should show data using only one physical port. Then bring your switch into the picture and re-test, bear in mind as you create EtherChannels on the hosts you need to replicate this on the switch.

If this is a single switch then your results should be good enough. If this is part of a bigger campus network there are many other factors that need to be taken into consideration, and this is a bit too complex to discuss on a forum like this.

Good luck

Andy