The ESX server NPM setting needs to be changed for VMAX.  You are probably hitting the IOPs limit on the FA CPU, not the throughput limit of the channel.

See VMWare KB article 2072070

http://kb.vmware.com/selfservice/microsites/microsite.do

Also please let us know what your IO size is.  You may be correct that going to 8gb won't give any more performance if you IO sizes are small (< 64K)

Ok, so large IO sizes.  You may find you get better throughput with IO sizes smaller than 512K blocks, such as 256K or 128K.  

The size of the data you are testing will fit entirely in cache, so you are only testing reads and writes to cache, no backend IO.

The change for the NMP should help the throughput in that case.

We are using 512KBytes block size, we perform sequential read for 30 seconds on a test file of 16GBytes, then 30 secs of random read, 30 secs of sequential write, 30 secs of random write, always on the same test file.

The result are quite stable and repeatable.

Yes we have jest applied the KB 2072070.

Ok 800MB/sec on a 8Gb link. So on a 4Gb link we have to reach 400MB/sec on a single FA port. You are saying that with a single host, on a single device we can't reach 400MB/sec? It's nonsense.

Different architecture in the VNX than VMAX.  We have dedicated CPUs for pairs of ports on the 40K.  Also we do one IO at a time for each device.  We need more active ports and devices to get full performance.  150MB/sec is about the maximum throughput for a single device on a single FA CPU.  I hope your LUN is a striped meta device, although the meta won't help as much for the sequential IO, as it is striped at 960K.

You achieved better performance using the other path management options, so I'm sure the round robin setting is what is slowing the IO for that test.

BTW, how do you have the active FA CPUs distributed across directors?  You should not have more than one port per FA CPU active.

The fact the test file fit entirely in cache is perfect, we are not interested in testing backend IO, we have to know WHY using a single FA port we achieve only 150MB/sec.

Same test on different storage (fare less expensive and powerful than VMAX) has given twice the throughput (in fixed path).

BTW, since all the data is in cache, where the data lives on the backend is not relevant.

And you may get better throughput for such a test from a lot of FC drives over EFDs.  For sequential large block reads and writes a FC drive is about the same as an EFD per spindle, and chances are you have a lot more FC spindles than EFD spindles.

And again, you may want to try different IO sizes, you may get better throughput with 64K, 128K or 256K.

How many active FA ports total?  And I hope none are sharing the same FA CPU.