[development][profile][dpdk] KK程序性能调优

KK程序:

1. 两个线程,第一个从DPDK收包,通过一个ring数据传递给第二个线程。第二个线程将数据写入共享内存。

2. 第二个内存在发现共享内存已满时,会直接丢弃数据。

3. 线程二有个选项debug,用于每一次ring_dequeue之后,都将数据写入内存。

  当这个选项为on时,内存未满,也不会丢包。

现象:当内存已满的时候,可以千兆线速收包。当内存未满时,丢包率为20%。

分别做三次gprof:

1. gmon-empty-off.txt

                0.08    0.42  471955/471955      kk_assemble_pool_packet_process [21]
[23]     1.8    0.08    0.42  471955         tcp_packet_process [23]
                0.02    0.10  739629/739629      _assemble_session_find [36]
                0.01    0.08  512494/512494      kk_tcp_session_request_find [43]
                0.01    0.07  665134/818327      kk_table_entries_timeout_free [39]
                0.04    0.04  330145/330145      _three_way_handshake_process [45]
                0.00    0.06  456472/461571      _tcp_data_assemble_process [48]

2. gmon-full-off.txt 

                0.08    0.48 3746819/3746819     kk_assemble_pool_packet_process [15]         
[16]     2.4    0.08    0.48 3746819         tcp_packet_process [16]                          
                0.09    0.08 3754332/3782242     kk_table_entries_timeout_free [26]           
                0.06    0.10 3746592/3746592     _assemble_session_find [27]                  
                0.01    0.09 3747382/3747382     kk_tcp_session_request_find [32]             
                0.03    0.01 3702531/3702531     _three_way_handshake_process [45]            
                0.00    0.00   56275/71247       _tcp_data_assemble_process [4295]

3. gmon-mid-on.txt

                0.10    0.55 3742005/3742005     kk_assemble_pool_packet_process [15]
[16]     2.3    0.10    0.55 3742005         tcp_packet_process [16]
                0.10    0.11 3745003/3745003     _assemble_session_find [21]
                0.06    0.09 3753439/3777518     kk_table_entries_timeout_free [25]
                0.02    0.11 3743276/3743276     kk_tcp_session_request_find [30]
                0.02    0.04 3689792/3689792     _three_way_handshake_process [45]
                0.00    0.00   64598/81267       _tcp_data_assemble_process [4295]

根据以上内容,对比一个关键步骤里的函数执行站比。可以发现。1中find查询的占比明确比其他两种情况更高。 而现象上也是情况1会有丢包,情况2,3不丢包。

再次测试,查看这三次的会话数。

1. gmon-empty-off.txt

name: tcp_assemble_task_1, size: 0, free: 1048575, pkts: 559700, session: 11688, hit: 0, drop: 0
name: udp_assemble_task_1, size: 0, free: 1048575, pkts: 0, session: 0, hit: 0, drop: 0
queue: 0, ipacket: 2595943, imissed: 1204057 self_counter: 2595943
queue: 0, total_tsc: 8825007304, tsc/pkt: 3399.538166

2. gmon-full-off.txt 

name: tcp_assemble_task_1, size: 0, free: 1048575, pkts: 48176, session: 16740, hit: 0, drop: 33152
name: udp_assemble_task_1, size: 0, free: 1048575, pkts: 0, session: 0, hit: 0, drop: 0
queue: 0, ipacket: 3800000, imissed: 0 self_counter: 3800000
queue: 0, total_tsc: 8785588132, tsc/pkt: 2311.996877

3. gmon-full-on.txt

name: tcp_assemble_task_1, size: 0, free: 1048575, pkts: 50746, session: 16982, hit: 0, drop: 33600
name: udp_assemble_task_1, size: 0, free: 1048575, pkts: 0, session: 0, hit: 0, drop: 0
queue: 0, ipacket: 3800000, imissed: 0 self_counter: 3800000
queue: 0, total_tsc: 8868949684, tsc/pkt: 2333.934127

并未发现规律。

使用新数据再次做次测试:

每15个包1个http会话。共270000个会话,按顺序组装,4050000个包。

1. empty_on

name: tcp_assemble_task_1, size: 0, free: 1048575, pkts: 1754014, session: 270001, hit: 0, drop: 528988
name: udp_assemble_task_1, size: 0, free: 1048575, pkts: 0, session: 0, hit: 0, drop: 0
queue: 0, ipacket: 4050000, imissed: 0 self_counter: 4050000
queue: 0, total_tsc: 17489586080, tsc/pkt: 4318.416316

2. empty_off

name: tcp_assemble_task_1, size: 0, free: 1048575, pkts: 2429992, session: 269999, hit: 0, drop: 0
name: udp_assemble_task_1, size: 0, free: 1048575, pkts: 0, session: 0, hit: 0, drop: 0
queue: 0, ipacket: 4050000, imissed: 0 self_counter: 4050000
queue: 0, total_tsc: 19613438800, tsc/pkt: 4842.824395

与上一组同样的测试数据,但是每5000个作为一组并发。

1. empty_on

name: tcp_assemble_task_1, size: 0, free: 1048575, pkts: 285000, session: 270000, hit: 0, drop: 540000
name: udp_assemble_task_1, size: 0, free: 1048575, pkts: 0, session: 0, hit: 0, drop: 0
queue: 0, ipacket: 4015852, imissed: 34148 self_counter: 4015852
queue: 0, total_tsc: 11696532776, tsc/pkt: 2912.590597

2. empty_off

name: tcp_assemble_task_1, size: 0, free: 1048575, pkts: 2068418, session: 235000, hit: 0, drop: 0
name: udp_assemble_task_1, size: 0, free: 1048575, pkts: 0, session: 0, hit: 0, drop: 0
queue: 0, ipacket: 3756940, imissed: 293060 self_counter: 3756940
queue: 0, total_tsc: 17565322544, tsc/pkt: 4675.433343

最后,是并发数的问题:

KK程序的最大并发数,只能处理到4000.

name: tcp_assemble_task_1, size: 0, free: 1048575, pkts: 2430000, session: 270000, hit: 0, drop: 0 tcp_session: 1 max_concurrent: 5000
name: udp_assemble_task_1, size: 0, free: 1048575, pkts: 0, session: 0, hit: 0, drop: 0 tcp_session: 1 max_concurrent: 5000
queue: 0, max_concurrent: 5000
queue: 0, ipacket: 4020940, imissed: 29060 self_counter: 4020940
queue: 0, total_tsc: 19906132788, tsc/pkt: 4950.616718
原文地址:https://www.cnblogs.com/hugetong/p/7251733.html