Date: Mon, 4 Dec 2023 11:52:21 +0800
From: Wen Gu <guwen@...ux.alibaba.com>
To: shaozhengchao <shaozhengchao@...wei.com>
Cc: "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
 yuehaibing <yuehaibing@...wei.com>, "Libin (Huawei)"
 <huawei.libin@...wei.com>, Dust Li <dust.li@...ux.alibaba.com>,
 tonylu_linux <tonylu@...ux.alibaba.com>, "D. Wythe"
 <alibuda@...ux.alibaba.com>
Subject: Re: [question] smc: how to enable SMC_LO feature


On 2023/12/4 11:22, shaozhengchao wrote:
> 
> 
> On 2023/11/23 14:15, shaozhengchao wrote:
>>
>>
>> On 2023/11/23 10:21, Wen Gu wrote:
>>>
>>>
>>> On 2023/11/21 20:14, shaozhengchao wrote:
>>>> Hi Wen Gu:
>>>> Currently, I am interested in the SMC_LOOPBACK feature proposed
>>>> by you. Therefore, I use your patchset[1] to test the SMC_LO feature on
>>>> my x86_64 environment and kernel is based on linux-next, commit: 5ba73bec5e7b.
>>>> The test result shows that the smc_lo feature cannot be enabled. Here's
>>>> my analysis:
>>>>
>>>> 1. Run the following command to perform the test, and then capture
>>>> packets on the lo device.
>>>> - serv:  smc_run taskset -c <cpu> sockperf sr --tcp
>>>> - clnt:  smc_run taskset -c <cpu> sockperf  tp --tcp --msg-size=64000 -i 127.0.0.1 -t 30
>>>>
>>>> 2. Use Wireshark to open packets. It is found that the VCE port replies with
>>>> SMC-R-Deline packets.
>>>> [cid:image001.png@...A1CB4.F1052C30]
>>>>
>>>> 3. Rx
>>>> When smc_listen_work invokes smc_listen_v2_check, the VCE port returns
>>>> a Decline packet because eid_cnt and flag.seid in the received packet are both 0.
>>>>
>>>> 4. Tx
>>>> In smc_clc_send_proposal,
>>>> v2_ext->hdr.eid_cnt = smc_clc_eid_table.ueid_cnt;
>>>> v2_ext->hdr.flag.seid = smc_clc_eid_table.seid_enabled;
>>>>
>>>> When smc_clc_init, ueid_cnt=0, and in the x86_64 environment, seid_enabled is
>>>> always equal to 0.
>>>>
>>>> So, I must call smc_clc_ueid_add function to increase ueid count?
>>>> But I don't see where operations can be added, may I missed something?
>>>>
>>>
>>> Hi Zhengchao Shao,
>>>
>>> Yes. When using SMC-D in non-s390 architecture (like x86 here), A common
>>> UEID should be set. It can be set by following steps:
>>>
>>> - Install smc-tools[1].
>>>
>>> - Run # smcd ueid add <ueid> in loopback test environment.
>>>
>>>    EID works as an ID to indicate the max communication space of SMC. When SEID is
>>>    unavailable, an UEID is required.
>>>
>> Hi Wen Gu:
>>      Thank you for your reply. This is very useful for me. And I will
>> be happy to learn from it.
>>
>> Thanks
>>
>> Zhengchao Shao
>>> - Then run the test.
>>>
>>> Hope this works for you :)
>>>
>>> [1] https://github.com/ibm-s390-linux/smc-tools
>>>
>>> Regards,
>>> Wen Gu
>>>
>>>> Could you give me some advice? Thanks very much.
>>>>
>>>> Zhengchao Shao
>>>>
>>>>
>>>> [1]link: 
>>>> https://patchwork.kernel.org/project/netdevbpf/cover/1695568613-125057-1-git-send-email-guwen@linux.alibaba.com/
>>>>
>>>
>>
>>
> Hi Wen Gu:
>      I have test as following, but the performance is really
> degraded. Now I have no idea.
> 1. add ueid
> run: smcd ueid add 16
> kernel message:
> [ 5252.009133] NET: Registered PF_SMC protocol family
> [ 5252.009233] smc: adding smcd device smc_lo with pnetid
> 2. start server
> smc_run taskset -c 1 sockperf sr --tcp
> 3. start client
> smc_run taskset -c 3 sockperf tp  --tcp --msg-size=64000 -i 127.0.0.1 -t 30
> 
> The test results are as follows:
>                TCP                  SMC-lo
> Bandwidth(MBps)         1890.56               1300.41(-31.22%)
> 
> I didn't find a better direction when I initially positioned it. No
> error is recorded in the kernel log, and the smcd statistics are normal.
> [root@...alhost smc-tools]# smcd stats
> SMC-D Connections Summary
>    Total connections handled             2
>    SMC connections                       2
>    Handshake errors                      0
>    Avg requests per SMC conn       1277462.0
>    TCP fallback                          0
> 
> RX Stats
>    Data transmitted (Bytes)    40907328000 (40.91G)
>    Total requests                  1277190
>    Buffer full                          45 (0.00%)
>              8KB    16KB    32KB    64KB   128KB   256KB   512KB  >512KB
>    Bufs        0       0       0       2       0       0       0       0
>    Reqs   638.0K       0       0  639.2K       0       0       0       0
> 
> TX Stats
>    Data transmitted (Bytes)    40907328000 (40.91G)
>    Total requests                  1277734
>    Buffer full                      638239 (49.95%)
>    Buffer full (remote)                  0 (0.00%)
>    Buffer too small                      0 (0.00%)
>    Buffer too small (remote)             0 (0.00%)
>              8KB    16KB    32KB    64KB   128KB   256KB   512KB  >512KB
>    Bufs        0       0       0       0       0       0       0       0
>    Reqs        0       0       0  1.278M       0       0       0       0
> 
> Extras
>    Special socket calls                  1
> 
> I captured the perf information and found that the percentage of
> rep_movs_alternative and _raw_spin_unlock_irqrestore functions was high
> during tx and rx.
> 36.12%  [kernel]         [k]rep_movs_alternative
> 14.23%  [kernel]         [k]_raw_spin_unlock_irqrestore
> 
> I've attached the flame map. Could you help analyze it? What I missed?
> Thanks.

Hi Zhengchao Shao,

Since sndbuf and RMB in SMC are pre-alloced ringbuf and won't grow dynamically
like TCP, it is necessary to appropriately increase the default value of smc
sk_sndbuf and sk_rcvbuf before testing throughput.

Set this and try again:

# sysctl -w net.smc.wmem=1048576
# sysctl -w net.smc.rmem=1048576

(The initial value of wmem and rmem are 64K)

Regards,
Wen Gu

> 
> Zhengchao Shao

Powered by blists - more mailing lists