Re: [PATCH net-next-2.6] net: speedup udp receive path

[Eric Dumazet <eric.dumazet@gmail.com>]

Le jeudi 29 avril 2010 à 20:12 +0800, Changli Gao a écrit :
> On Thu, Apr 29, 2010 at 7:35 PM, jamal <hadi@cyberus.ca> wrote:
> >
> > Same here - even in my worst case scenario 88.5% of 750Kpps > 600Kpps.
> > Attached is history results to make more sense of what i am saying:
> > we have net-next kernels from apr14, apr23, apr23 with changlis change,
> > apr28, apr28 with your change. What you'll see is non-rps (blue) gets
> > better and rps (Orange) gets better slowly then by apr28 it is worse.
> 
> Did the number of IPIs increase in the apr28 test? The finial patch
> with Eric's change may introduce more IPIs. And I am wondering why
> 23rdcl-non-rps is better than before. Maybe it is the side effect of
> my patch: enlarge the netdev_max_backlog.
> 
> 

Changli, I wonder how you can cook "performance" patches without testing
them at all for real... This cannot be true ?

When the cpu doing the device softirq is flooded, it handles 300 packets
per net_rx_action() round (netdev_budget), so sends at most 6 ipis per
300 packets, with or without my patch, with or without your patch as
well.

(At most because if remote cpus are flooded as well, they dont
napi_complete so no IPI needed at all)

(My patch had an effect only on normal load, ie one packet received in a
while... up to 50.000 pps I would say). And it also has a nice effect on
non RPS loads (mostly the more typical load for following years).
If a second packet comes 3us after the first one, and before 2nd CPU
handled it, we _can_ afford an extra IPI.

750.000/50 = 15.000 IPI per second.

Even with 200.000 IPI per second, 'perf top -C CPU_IPI_sender' shows
that sending IPI is very cheap (maybe ~1% of cpu cycles)

# Samples: 32033467127
#
# Overhead         Command      Shared Object  Symbol
# ........  ..............  .................  ......
#
    18.05%            init  [kernel.kallsyms]  [k] poll_idle
    10.91%            init  [kernel.kallsyms]  [k] bnx2x_rx_int
    10.42%            init  [kernel.kallsyms]  [k] eth_type_trans
     5.72%            init  [kernel.kallsyms]  [k] kmem_cache_alloc_node
     5.43%            init  [kernel.kallsyms]  [k] __memset
     5.20%            init  [kernel.kallsyms]  [k] get_rps_cpu
     4.82%            init  [kernel.kallsyms]  [k] __slab_alloc
     4.34%            init  [kernel.kallsyms]  [k] get_partial_node
     4.22%            init  [kernel.kallsyms]  [k] _raw_spin_lock
     3.41%            init  [kernel.kallsyms]  [k] __kmalloc_node_track_caller
     3.01%            init  [kernel.kallsyms]  [k] __alloc_skb
     2.22%            init  [kernel.kallsyms]  [k] enqueue_to_backlog
     2.10%            init  [kernel.kallsyms]  [k] vlan_gro_common
     1.34%            init  [kernel.kallsyms]  [k] swiotlb_map_page
     1.25%            init  [kernel.kallsyms]  [k] skb_put
     1.06%            init  [kernel.kallsyms]  [k] _raw_spin_lock_irqsave
     0.92%            init  [kernel.kallsyms]  [k] dev_gro_receive
     0.88%            init  [kernel.kallsyms]  [k] swiotlb_dma_mapping_error
     0.83%            init  [kernel.kallsyms]  [k] vlan_gro_receive
     0.83%            init  [kernel.kallsyms]  [k] __phys_addr
     0.83%            init  [kernel.kallsyms]  [k] __napi_complete
     0.83%            init  [kernel.kallsyms]  [k] default_send_IPI_mask_sequence_phys
     0.77%            init  [kernel.kallsyms]  [k] is_swiotlb_buffer
     0.76%            init  [kernel.kallsyms]  [k] __netdev_alloc_skb
     0.74%            init  [kernel.kallsyms]  [k] deactivate_slab
     0.73%            init  [kernel.kallsyms]  [k] netif_receive_skb
     0.72%            init  [kernel.kallsyms]  [k] unmap_single
     0.69%            init  [kernel.kallsyms]  [k] csd_lock
     0.63%            init  [kernel.kallsyms]  [k] bnx2x_poll
     0.61%            init  [kernel.kallsyms]  [k] bnx2x_msix_fp_int
     0.59%            init  [kernel.kallsyms]  [k] irq_entries_start
     0.59%            init  [kernel.kallsyms]  [k] swiotlb_sync_single
     0.54%            init  [kernel.kallsyms]  [k] get_slab
     0.46%            init  [kernel.kallsyms]  [k] napi_skb_finish