[Info-vax] Linux 40 GbE and 100 GbE NIC Performance

[terry+googleblog at tmk.com]

On Friday, January 23, 2015 at 4:52:17 PM UTC-5, Stephen Hoffman wrote:
> These same difficulties and performance requirements will apply to VMS 
> network drivers, if (when?) support for faster NICs beyond the current 
> 10 GbE NIC options is added to VMS.   Whether VMS can respond as 
> quickly as needed, and whether there might be a need to dedicate one or 
> more cores for the NIC processing?

I don't have anything faster than 10GbE here at home, but I have a reasonable number of 10GbE hosts (10 or so). 10GbE and beyond is going to be used mainly for bulk transfers. I have done a reasonable amount of benchmarking and I can get "wire speed" between two hosts using iperf:

(0:57) gate:/sysprog/terry# iperf -c axpemu
------------------------------------------------------------
Client connecting to axpemu, TCP port 5001
TCP window size: 32.0 KByte (default)
------------------------------------------------------------
[  3] local 10.20.30.40 port 64332 connected with 10.20.30.62 port 5001
[ ID] Interval       Transfer     Bandwidth
[  3]  0.0-10.0 sec  11.5 GBytes  9.88 Gbits/sec

The systems are both Dell PowerEdge R710's. axpemu is the one I'm running the Alpha emulator on, and is a quad-core 3GHz system (nothing special, in fact somewhat slow by today's standards).

I could not find an implementation of iperf for OpenVMS Alpha. Of course, even if I had it, there would be a second level of indirection through the emulator.

The iperf results are the absolute best case, as the data doesn't touch the disk or have much done to it other than transmitting / receiving it. It does show that the FreeBSD kernel, TCP stack and user-mode applications are capable of wire-speed operation, even with single-thread operation.

I have been able to get real-world transfers in the 650Mbyte/sec range between one of my RAIDzilla 2 file servers and a PC running Windows 7 with a SSD as the storage device.