[PATCH net-next v3 0/4] net: mvneta: improve rx/tx performance

Hi Gregory,

On Mon, 20 Feb 2017 15:21:35 +0100 Gregory CLEMENT wrote:

Hi Jisheng,
 
 On lun., f?vr. 20 2017, Jisheng Zhang [off-list ref] wrote:

quoted

In hot code path such as mvneta_rx_swbm(), we access fields of rx_desc
and tx_desc. These DMA descs are allocated by dma_alloc_coherent, they
are uncacheable if the device isn't cache coherent, reading from
uncached memory is fairly slow.

patch1 reuses the read out status to getting status field of rx_desc
again.

patch2 avoids getting buf_phys_addr from rx_desc again in
mvneta_rx_hwbm by reusing the phys_addr variable.

patch3 avoids reading from tx_desc as much as possible by store what
we need in local variable.

We get the following performance data on Marvell BG4CT Platforms
(tested with iperf):

before the patch:
sending 1GB in mvneta_tx()(disabled TSO) costs 793553760ns

after the patch:
sending 1GB in mvneta_tx()(disabled TSO) costs 719953800ns

we saved 9.2% time.

patch4 uses cacheable memory to store the rx buffer DMA address.

We get the following performance data on Marvell BG4CT Platforms
(tested with iperf):

before the patch:
recving 1GB in mvneta_rx_swbm() costs 1492659600 ns

after the patch:
recving 1GB in mvneta_rx_swbm() costs 1421565640 ns  

Could you explain who you get this number?

Thanks for your review.

The measurement is simple: record how much time we spent in mvneta_rx_swbm()
for receiving 1GB data, something as below:

mvneta_rx_swbm()
{
	static u64 total_time;
	u64 t1, t2;
	static u64 count;

	t1 = sched_clock();
	...

	if (rcvd_pkts) {
		...
		t2 = sched_clock() - t1;
		total_time += t2;
		count += rcvd_bytes;;
		if (count >= 0x40000000) {
			printk("!!!! %lld %lld\n", total_time, count);
			total_time = 0;
			count = 0;
		}
	...
}

receiving 1GB in 1.42 second means having a bandwidth of
8/1.42=5.63 Gb/s, that means that you are using at least a 10Gb
interface.

hmmm, we just measured the time spent in mvneta_rx_swbm(), so we can't solve
the bandwidth as 8/1.42, what do you think?

When I used iperf I didn't have this kind of granularity:
iperf -c 192.168.10.1 -n 1024M
------------------------------------------------------------
Client connecting to 192.168.10.19, TCP port 5001
TCP window size: 43.8 KByte (default)
------------------------------------------------------------
[  3] local 192.168.10.28 port 53086 connected with 192.168.10.1 port 5001
[ ID] Interval       Transfer     Bandwidth
[  3]  0.0- 9.1 sec  1.00 GBytes   942 Mbits/sec

Also without HWBM enabled (so with the same configuration of your test),
I didn't noticed any improvement with the patch set applied. But at

From bandwidth point of view, yes, there's no improvement. But from cpu

time/load point of view, I do see a trivial improvement. Could you also
did a simple test from your side to see whether we have similar improvement
data?

Thanks,
Jisheng

least I didn't see any regression with or without HWBM.

Gregory

quoted

We saved 4.76% time.

Basically, patch1 and patch4 do what Arnd mentioned in [1].

Hi Arnd,

I added "Suggested-by you" tag, I hope you don't mind ;)

Thanks

[1] https://www.spinics.net/lists/netdev/msg405889.html

Since v2:
  - add Gregory's ack to patch1
  - only get rx buffer DMA address from cacheable memory for mvneta_rx_swbm()
  - add patch 2 to read rx_desc->buf_phys_addr once in mvneta_rx_hwbm()
  - add patch 3 to avoid reading from tx_desc as much as possible

Since v1:
  - correct the performance data typo


Jisheng Zhang (4):
  net: mvneta: avoid getting status from rx_desc as much as possible
  net: mvneta: avoid getting buf_phys_addr from rx_desc again
  net: mvneta: avoid reading from tx_desc as much as possible
  net: mvneta: Use cacheable memory to store the rx buffer DMA address

 drivers/net/ethernet/marvell/mvneta.c | 80 +++++++++++++++++++----------------
 1 file changed, 43 insertions(+), 37 deletions(-)

-- 
2.11.0