3PAR Iometer Ninja Stars IOPS disrepancy

[cqcmbr]

Hello,

The problem is I am unable to get near the provided IOPS in the latest Ninja Stars tool. I get around half the expected bandwith and half the IOPS with IOMETER (16 workers, 8 outstanding IO`s, full disk). Increasing / Decreasing the Manager or Worker count doesnt change much up to the point when only latency increases.

For example, Full Write with 64K blocks results in half the expected results.

If it helps, SSMC shows that the host is 100% busy when IOMETER is run.

Do you have any advice on how to get better IOMETER results?

Has anyone managed to replicate Ninja Stars proposed results?

Could putting the FC card into a x16 PCIE slot improve performance x2?

THE SETUP:

3PAR 8400 storage array for testing, 400GB SSD disks in RAID 5, 3+1.

The host that the 100GB volume is exported to is a DL380 G9 with Windows 2012 R2 (No Hypervisor to reduce overhead)

No switches, direct attach. DL380 has a SN1000Q 16Gb FC card in a x8 riser slot. The 3Par ports are the standart 16G 0:0:1 and 1:0:1 ports. Latest firmware and drivers for everything. Round robin in MPIO, FC cables are in good condition.

-In addition, ESXi with Round Robin gets fairly worse results.

Thank you!

TL:DR- HOW TO GET PROPER IOMETER RESULTS?

!!! UPDATE!!! - I added an additional SN1000Q 16Gb card, and the results almost doubled. So my guess is that the bottleneck is definitely with the windows host server. And it has something to do with throughput per port. Any ideas gents?

[Cali]

I tested also a 8400 AFA with IO Meter.

We need 6 Windows Hosts (8G FC) to satturate the Storage.

Every Host have ESX with 2x Windows VM IO Meter and RAW Disks.

After the 7. Host we get no more IOPS out of the Storage.

Every Storage SFP have own Buffers and CPU Core (some changes in 3.3.1), so using additional SFPs give more Performance.

Do you use the second SAS Ports (DP-1) by adding a Shelf?