On Wed 18-11-15 16:48:22, Johannes Weiner wrote:
[...]

So I ran perf record -g -a netperf -t TCP_STREAM multiple times inside
a memory-controlled cgroup, but mostly mem_cgroup_charge_skmem() does
not show up in the profile at all. Once it was there with 0.00%.

OK, this sounds very good! This means that most workloads which are not
focusing solely on the network traffic shouldn't even notice. I can
imagine that workloads with high throughput demands would notice but I
would also expect them to disable the feature.

Could you add this information to the changelog, please?

I ran another test that downloads the latest kernel image from
kernel.org at 13MB/s (on my i5 laptop) and it looks like this:

     0.02%     0.01%  irq/44-iwlwifi   [kernel.kallsyms]           [k] mem_cgroup_charge_skmem
             |
             ---mem_cgroup_charge_skmem
                __sk_mem_schedule
                tcp_try_rmem_schedule
                tcp_data_queue
                tcp_rcv_established
                tcp_v4_do_rcv
                tcp_v4_rcv
                ip_local_deliver
                ip_rcv
                __netif_receive_skb_core
                __netif_receive_skb
                netif_receive_skb_internal
                napi_gro_complete

The runs vary too much for this to be measurable in elapsed time.

-- 
Michal Hocko
SUSE Labs