dnsperf: -Q max_qps limit does not work as advertised

Version affected: 2.10.0

First, definition of -Q in the manual page:

-Q max_qps Limits the number of requests per second. There is no default limit.

(Emphasis added by me.)

This does not match the real behavior of that feature. Right now the code looks like this:

        /* Rate limiting */
        if (tinfo->max_qps > 0) {
            run_time = now - times->start_time;
            req_time = (MILLION * stats->num_sent) / tinfo->max_qps;
            if (req_time > run_time) {
                usleep(req_time - run_time);
                now = perf_get_time();
                continue;
            }
        }

It calculates when the next packet should be sent based on total number of packets, not number of packets in given second.

Consequently, if the server is slower than specified -Q for a period of time, unsent packets “accumulate” and then the -Q value is overshot, possibly very significantly.

Here is example of such test run in table. Columns 2, 3, 4 are number of packets as reported by dnsperf. See time 25 s and following.

time [s]	sent	recvd	timeouts
1	100021	99673	0
2	100031	100024	0
3	100028	100027	0
4	100027	100031	0
5	100028	100030	0
6	100023	100023	323
7	100022	100023	0
8	100018	100015	0
9	100025	100020	0
10	100019	100022	0
11	100130	100134	0
12	100026	100026	0
13	100028	100033	0
14	100018	100018	0
15	100032	100025	0
16	100023	100027	0
17	100026	100017	0
18	100022	100028	0
19	100025	100020	0
20	100026	100025	0
21	100034	100039	0
22	100026	100029	0
23	100028	100014	0
24	100023	100040	0
25	100028	100023	0
26	51631	50659	0
27	0	0	0
28	0	0	0
29	0	0	0
30	0	0	0
31	9383	8383	1000
32	0	0	0
33	0	0	0
34	0	0	0
35	0	0	0
36	9388	8388	1000
37	0	0	0
38	0	0	0
39	0	0	0
40	0	0	0
41	9349	8350	1000
42	9	8	0
43	0	0	0
44	0	0	0
45	0	0	0
46	98089	98080	999
47	215355	215341	1
48	214683	214645	0
49	214870	214909	0
50	214906	214837	0
51	215887	215965	0
52	219043	218387	0
53	209458	209342	0
54	213390	213553	0
55	215100	215093	0
56	214460	214249	0
57	214883	215026	609
58	213354	213398	0
59	212175	211942	0

Simplest reproducer is to run dnsperf with -Q against an echo server, kill the server in the middle of test, and then restart it. dnsperf will send more queries as it tries to “catch up”.

$ yes '. A' | dnsperf -O suppress=timeout -t 1 -Q 100000 -l 10 -s 127.0.0.1 -S1 -p 5353
DNS Performance Testing Tool
Version 2.10.0

[Status] Command line: dnsperf -O suppress=timeout -t 1 -Q 100000 -l 10 -s 127.0.0.1 -S1 -p 5353
[Status] Sending queries (to 127.0.0.1:5353)
[Status] Started at: Thu Feb  2 16:12:30 2023
[Status] Stopping after 10.000000 seconds
1675350751.204563: 99992.803613
1675350752.205596: 100005.694118
1675350753.206756: 34156.378601
1675350754.207943: 0.000000
1675350755.209119: 0.000000
1675350756.210197: 0.000000
1675350757.211227: 226568.634307
1675350758.211343: 324562.350767
1675350759.211482: 114571.074621
[Status] Testing complete (time limit)

Statistics:

  Queries sent:         999995
  Queries completed:    999595 (99.96%)
  Queries lost:         400 (0.04%)

  Response codes:       NOERROR 999595 (100.00%)
  Average packet size:  request 17, response 17
  Run time (s):         10.000020
  Queries per second:   99959.300081

  Average Latency (s):  0.000160 (min 0.000005, max 0.000604)
  Latency StdDev (s):   0.000142