Re: [RFC] performance regression with commit-id<adb03115f459> ("net: get rid of an signed integer overflow in ip_idents_reserve()")

[Eric Dumazet]

On 7/24/19 10:38 AM, Zhangshaokun wrote:
> Hi,
> 
> I've observed an significant performance regression with the following commit-id <adb03115f459>
> ("net: get rid of an signed integer overflow in ip_idents_reserve()").

Yes this UBSAN false positive has been painful



> 
> Here are my test scenes:
> ----Server----
> Cmd: iperf3 -s xxx.xxx.xxxx.xxx -p 10000 -i 0 -A 0
> Kenel: 4.19.34
> Server number: 32
> Port: 10000 â 10032
> CPU affinity: 0 â 32
> CPU architecture: aarch64
> NUMA node0 CPU(s): 0-23
> NUMA node1 CPU(s): 24-47
> 
> ----Client----
> Cmd: iperf3 -u -c xxx.xxx.xxxx.xxx -p 10000 -l 16 -b 0 -t 0 -i 0 -A 8
> Kenel: 4.19.34
> Client number: 32
> Port: 10000 â 10032
> CPU affinity: 0 â 32
> CPU architecture: aarch64
> NUMA node0 CPU(s): 0-23
> NUMA node1 CPU(s): 24-47
> 
> Firstly, With patch <adb03115f459> ("net: get rid of an signed integer overflow in ip_idents_reserve()") ,
> clientâs cpu is 100%, and function ip_idents_reserve() cpu usage is very high, but the result is not good.
> 03:08:32 AM     IFACE   rxpck/s   txpck/s    rxkB/s    txkB/s   rxcmp/s   txcmp/s  rxmcst/s   %ifutil
> 03:08:33 AM      eth0      0.00      0.00      0.00      0.00      0.00      0.00      0.00      0.00
> 03:08:33 AM      eth1      0.00 3461296.00      0.00 196049.97      0.00      0.00      0.00      0.00
> 03:08:33 AM        lo      0.00      0.00      0.00      0.00      0.00      0.00      0.00      0.00
> 
> Secondly, revert that patch, use atomic_add_return() instead, the result is better, as below:
> 03:23:24 AM     IFACE   rxpck/s   txpck/s    rxkB/s    txkB/s   rxcmp/s   txcmp/s  rxmcst/s   %ifutil
> 03:23:25 AM        lo      0.00      0.00      0.00      0.00      0.00      0.00      0.00      0.00
> 03:23:25 AM      eth1      0.00 12834590.00      0.00 726959.20      0.00      0.00      0.00      0.00
> 03:23:25 AM      eth0      7.00     11.00      0.40      2.95      0.00      0.00      0.00      0.00
> 
> Thirdly, atomic is not used in ip_idents_reserve() completely ,while each cpu core allocates its own ID segment,
> Such as: cpu core0 allocate ID 0 â 1023, cpu core1 allocate 1024 â 2047, â,etc
> the result is the best:

Not sure what you mean.

Less entropy in IPv4 ID is not going to help when fragments _are_ needed.

Send 40,000 datagrams of 2000 bytes each, add delays, reorders, and boom, most of the packets will be lost.

This is not because your use case does not need proper IP ID that we can mess with them.

If you need to send packets very fast,  maybe use AF_PACKET ?

> 03:27:06 AM     IFACE   rxpck/s   txpck/s    rxkB/s    txkB/s   rxcmp/s   txcmp/s  rxmcst/s   %ifutil
> 03:27:07 AM        lo      0.00      0.00      0.00      0.00      0.00      0.00      0.00      0.00
> 03:27:07 AM      eth1      0.00 14275505.00      0.00 808573.53      0.00      0.00      0.00      0.00
> 03:27:07 AM      eth0      0.00      2.00      0.00      0.18      0.00      0.00      0.00      0.00
> 
> Because atomic operation performance is bottleneck when cpu core number increase, Can we revert the patch or
> use ID segment for each cpu core instead?


This has been discussed in the past.

https://lore.kernel.org/lkml/b0160f4b-b996-b0ee-405a-3d5f1866272e@xxxxxxxxx/

We can revert now UBSAN has been fixed.

Or even use Peter patch : https://lore.kernel.org/lkml/20181101172739.GA3196@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/

However, you will still hit badly a shared cache line, not matter what.

Some arches are known to have terrible LL/SC implementation :/