Re: Performance issues with Intel Fortville (XL710/ixl(4))

[Lakshmi Narasimhan Sundararajan]

Hi FreeBSD Team!

We seem to have found a problem to Tx performance. 

We found that the tx handling is spread on all CPUs causing probably cache trashing resulting in poor performance. 

But once we used cpuset to bind interrupt thread and iperf process to the same CPU, performance was close to line rate. I used userland cpuset command to perform this manually. I want this constrained in the kernel config/code through some tunables, and I am seeking your help/pointers in that regard.


My followup questions are as follows.

a) How are Tx interrupts steered from the NIC to the CPU on the transmit path? Would tx_complete# interrupt for packets transmitted from CPU#x, be serviced on the same CPU? If not, how to get this binding done?


b) I would like to use a pool of CPUs dedicated to service NIC interrupts. Especially on the transmit path, I would want the tx_interrupts to be handled on the same CPU on which request was submitted. How to get this done?


I played with the current ISR setting but I did not see any difference in how Interrupts are scheduled across CPU. The max interrupt threads even though set to one, the interrupt threads are scheduled on any CPU. Even if I set bindthreads to ‘1’. There is no difference in interrupt thread scheduling.


root at mau-da-27-4-1:~ # sysctl net.isr
net.isr.dispatch: direct
net.isr.maxthreads: 1
net.isr.bindthreads: 0
net.isr.maxqlimit: 10240
net.isr.defaultqlimit: 256
net.isr.maxprot: 16
net.isr.numthreads: 1


I would sincerely appreciate if you can provide some pointers on these items above.




Thanks

LN







From: Pokala, Ravi
Sent: ‎Wednesday‎, ‎May‎ ‎20‎, ‎2015 ‎3‎:‎34‎ ‎AM
To: freebsd-net at freebsd.org, jfv at freebsd.org, erj at freebsd.org
Cc: freebsd-hackers at freebsd.org, Lewis, Fred, Sundararajan, Lakshmi





Hi folks,

At Panasas, we are working with the Intel XL710 40G NIC (aka Fortville),
and we're seeing some performance issues w/ 11-CURRENT (r282653).

    Motherboard: Intel S2600KP (aka Kennedy Pass)
    CPU: E5-2660 v3 @ 2.6GHz (aka Haswell Xeon)
        (1 socket x 10 physical cores x 2 SMT threads) = 20 logical cores
    NIC: Intel XL710, 2x40Gbps QSFP, configured in 4x10Gbps mode
    RAM: 4x 16GB DDR4 DIMMs

What we've seen so far:

  - TX performance is pretty consistently lower than RX performance. All
numbers below are for unidrectional tests using `iperf':
        10Gbps links    threads/link    TX Gbps     RX Gbps     TX/RX
        1               1               9.02        9.85        91.57%
        1               8               8.49        9.91        85.67%
        1               16              7.00        9.91        70.63%
        1               32              6.68        9.92        67.40%

  - With multiple active links, both TX and RX performance suffer greatly;
the aggregate bandwidth tops out at about a third of the theoretical
40Gbps implied by 4x 10Gbps.
        10Gbps links    threads/link    TX Gbps     RX Gbps     % of 40Gbps
        4               1               13.39       13.38       33.4%

  - Multi-link bidirectional throughput is absolutely terrible; the
aggregate is less than a tenth of the theoretical 40Gbps.
        10Gbps links    threads/link    TX Gbps     RX Gbps     % of 40Gbps
        4               1               3.83        2.96        9.6% / 7.4%

  - Occasional interrupt storm messages are seen from the IRQs associated
with the NICs. Since that can impact performance, those runs were not
included in the data listed above.

Our questions:

  - How stable is ixl(4) in -CURRENT? By that, we mean both how quickly is
the driver changing, and does the driver cause any system instability?

  - What type of performance have others been getting w/ Fortville? In
40Gbps mode? In 4x10Gbps mode?

  - Does anyone have any tuning parameters they can recommend for this
card?

  - We did our testing w/ 11-CURRENT, but we will initially ship Fortville
running on 10.1-RELEASE or 10.2-RELEASE. The presence of RSS - even though
it is disabled by default - makes the driver back-port non-trivial. Is
there an estimate on when the 11-CURRENT version of the driver (1.4.1)
will get MFCed to 10-STABLE?

My colleagues Lakshmi and Fred (CCed) are working on this; please make
sure to include them if you have any comments.

Thanks,

Ravi