EVA4400/6400/8400 - more than 4gbps throughput per LUN possible?

[sam bell]

Hi,

I just have a short question: Am I correct in assuming that with the 4gbps EVAs named in the subject it is not possible to achieve more than 4gbps throughput per LUN even if a second path through the non-owning controller is available?

If I understand correctly, a LUN is presented over one controller only and cannot be accessed over both controllers at the same time (well, virtually this is possible but actually everything is proxied to the LUN owning controller so that actually only one controller can issue I/O to the LUN at any given time which limits the throughput to 4gbps).

Is that correct or am I wrong?

[Mattias Sjödin]

Hi,

You are correct in your assumption as far as I know. The LUN can be accessed through the non-owning controller, but the data will still be brought through the internal mirror port between the controllers.

Depending on the LUN preferred path setting, the controller pair will try to put the ownership on the controller where most of the host data is being transferred.

Hope that helped!

/m

[Johan Guldmyr]

Yes, it will change ownership. If there are more than 63% (or something like that 2/3 I guess) reads through one controller it will move it over to the other controller.

The recommendation is to spread the load between the controllers by assigning LUNs accordingly to each controller.

--

But, doesn't each controller have at least two ports? What if you connect to the 1 LUN via those 2 FC host ports? Wouldn't you then get 8Gbps? I would assume so?

--

On the EVA8400/6400 each controller has 4 host ports which makes it easier. As you can then have two ports per controller per fabric. If you take both host ports on an EVA4400 controller then you lose a bit of redundancy. But the EVA4400 comes with an option to have the embedded SAN switch (8Gbps) which increases the amount of connections. Note that the internal connection from the EVA4400-controller to the embedded SAN switch is still 4Gbps.

[sam bell]

Thanks for your replies. But is it really through that each ports on a controller delivers 4gbps so you could theoretically reach 16gbps with an EVA6400/8400 if all four ports are connected to the same fabric?

Or are the ports sharing the 4gbps of the controller and the ports are just for redundancy?

[sam bell]

Sorry I meant "true".

[Johan Guldmyr · Accepted solution]

Hi, well each port is 4gbps.

But, I don't know for sure though if you put all 4 on the same fabric if you can reach 16gbps speed (maybe you reach a limitation on something else like amount of disks).

[sam bell]

Thanks, I just confirmed this by testing the following:

- EVA 6400 with 64 disks
- 2 fabrics from which each is connected to 2 ports of each controller
- Host with two 2gbps FCAa, one connected to fabric A and one connected to fabric B

Running a sequential read benchmark with IOmeter and both FCA showed 380mb/s.

A second test with only one FCA enabled showed only 196mb/s.

So in theory, if I had two 8gbps adapters installed in the host, I should really be able to reach 16gbps per LUN (two paths to each controller per FCA == 2x 4gbps to controller A on FCA 1 and 2x 4gbps to controller B on FCA 2 == total 16gbps).

[sam bell]

The above is wrong, of course I meant:

2x 4gbps to controller A on FCA 1 == 8gbps
2x 4gbps to controller A on FCA 2 == 8gbps

...which results in theoretically possible 16gbps to each controller.

[Johan Guldmyr]

Yeup!

Great that you have the ability to test it right away!

The test you ran, did you have a vdisk/lun to test it on?
What kind of disks do you have and what kind of vraid did you have?

[sam bell]

I've run the test on a 10GB Vraid1 Vdisk; the disks are 15k 300GB FC disks.

The IOmeter test I've used was the one attached which is used by the VMware communities to compare storage systems.