em network issues

[Scott Long]

John Polstra wrote:
> On 19-Oct-2006 Scott Long wrote:
>> The performance measurements that Andre and I did early this year showed
>> that the INTR_FAST handler provided a very large benefit.
> 
> I'm trying to understand why that's the case.  Is it because an
> INTR_FAST interrupt doesn't have to be masked and unmasked in the
> APIC?  I can't see any other reason for much of a performance
> difference in that driver.  With or without INTR_FAST, you've got
> the bulk of the work being done in a background thread -- either the
> ithread or the taskqueue thread.  It's not clear to me that it's any
> cheaper to run a task than it is to run an ithread.
> 
> A difference might show up if you had two or more em devices sharing
> the same IRQ.  Then they'd share one ithread, but would each get their
> own taskqueue thread.  But sharing an IRQ among multiple gigabit NICs
> would be avoided by anyone who cared about performance, so it's not a
> very interesting case.  Besides, when you first committed this
> stuff, INTR_FAST interrupts were not sharable.

Interrupt sharing is one consideration.  And sometimes sharing can't be
avoided; I'm working with a product that has 2 dual port e1000 cards,
4 dual port FC cards, and an onboard dual port e1000 chip and dual port
SCSI chip.  All of those devices are only on 2 PCI buses.  So, sharing
is inevitable.  MSI might help mitigate this, once it becomes a reality
in FreeBSD.

The cost of the APIC operation can be a factor.  There is a single
spinlock for all APICs, so you can get contention with multiple CPUs
servicing interrupts and colliding on the APIC lock.  When there isn't
much contention, the cost of the APIC spinlock is no more than the cost
of the taskqueue lock, and the scheduling locks are about the same.

> 
> Another change you made in the same commit (if_em.c revision 1.98)
> greatly reduced the number of PCI writes made to the RX ring consumer
> pointer register.  That would yield a significant performance
> improvement.  Did you see gains from INTR_FAST even without this
> independent change?

PCI writes are store-and-forward so they are practically free, unlike
PCI reads.  There was speculation that interrupting the chip less with
fewer writes might be a benefit, but I don't think that this was ever
tested by itself.

The big win came from moving the locking outside of the basic interrupt
handler.  The handler can now run completely free of any other driver or
stack contention.  False interrupts (especially for the shared case) can
be handled without blocking the rest of the driver.  Since there is no
ithread sharing, there is also a fairly deterministic amount of time to
get to the interrupt handler now; so the chance of rx  overflows goes
down.  Even if one thread is stuck in a tight loop doing tx or
tx-complete operations, the taskqueue can run the rx path without any
contention.  Also, in the non-shared case, the amount of time needed to
do the register read in the handler no longer delays the upper or lower
half of the driver.

Andre measured about a 20% gain with just this, IIRC.  I then took the 
next step and pushed locking out of the RX path.  That dramatically 
increased forwarding performance, as the rx path of one driver could
drive the tx path of the other driver without any contention from either
driver.  We came close to getting 1mpps in the fast-forward case, and
we suspect that the only reason that we didn't reach the 1m mark is
because prefetching was turned off accidentally on the chip.

This last change brought the overall performance gain to about 50% more
than the unmodified driver.  I certainly don't attribute that all to the
INTR_FAST change, but like I stated above, a good portion of it is.

Note that I started the work on if_em not to improve performance, but to
prototype a way to eliminate the interrupt aliasing that was happening
on the Intel chipsets.  I only asked Andre and others to test it for
performance because I didn't want it to be a pessimization.  Once we
discovered that it helped, I went the extra steps with optimizing the
rx path.

Scott