Enable ssl_poll_perf test to run as client/server. #63

Sashan · 2025-11-05T07:17:58Z

This enables as to evaluate performance of OpenSSL QUIC stack against other QUIC implementations.

Sashan · 2025-11-05T11:10:50Z

those changes were tested with tqclient/tqserver:

# git clone --recurse https://github.com/Tencent/tquic
# cd tquic
# cargo build --all # to build also tools/examples
# echo $?

if build is successful tqserver can be launched:

# cd target/debug/
# ./tquic_server -l 127.0.0.1:7070 -c path/to/servercert.pem -k path/to/serverkey.pem
# echo 'payload' > payload

Feel free to use servercert.pem and serverkey.pem found in $OPENSSL_SRCDIR/test/certs.
Command above starts QUIC server listening to port 7070. Use command here to connect:

# cd $PERFTOOLS./source/build
# ./ssl_poll_perf -m c -a 0 -c 100 -b 1 -u 0 http://127.0.0.1:7070/payload

command above launches client which performs 100 connections to tquic_server. Each connection
opens single bi-directional stream and fetches file payload.

to run tquic_client against ssl_poll_perf server, launch server first:

# cd $PERFTOOLS./source/build
# ./ssl_poll_perf -p 8080  -m s -a 0 path/to/servercert.pem path/to/serverkey.pem

To download 7 bytes from server above use URL http://127.0.0.1:8080/7 using
tquic_client command below:

./tquic_client \
    -t 1 \ # use one thread
    --max-concurrent-conns 100 \ # 100 concurrent connections
    --max-concurrent-requests 100 \ # 100 concurrent requests
    --max-requests-per-conn 1 \ # one request per connection
    --total-requests-per-thread 100 \ # 100 requests
    -d 0 # wait until all requests are completed
    -a http/0.9 # use http/0.9
    http://127.0.0.1:8080/7

quarckster · 2025-11-05T13:05:00Z

I'm wondering how to better run this in our performance benchmark lab. Is it enough to run both server and client on the same machine? Or the client and the server must be resided on separate hosts. In that case does it matter the platform of the counterpart? It will also bring questions of the simultaneous access to the counterpart from different testing machines. How will it affect on the results?

jogme

First iteration

jogme · 2025-11-05T08:46:21Z

README.md

 -w - the size of request body, the maximum size is 100MB. The default size is 64.
 -p - port number to use
 -t - terse output
+-m - mode {client|server}, self-test mode when ommitted


nit: mention somehow the one character option too (c,s)

jogme · 2025-11-05T14:10:35Z

source/ssl_poll_perf.c

+    time_t t;
+
+    t = time(&t);
+    ctime_r(&t, date_str);


nit: check on return code

jogme · 2025-11-05T14:15:28Z

source/ssl_poll_perf.c

+}
+
+static struct rr_buffer *
+setup_response_0_9(struct rr_txt_full *rtf, unsigned int fsize)


why do we need HTTP 0.9?

the tqserver/tqclient do not understand http/1.0. however the http/0.9 seems to be supported by various QUIC stack testing tools.

Though I'm not sure about keeping support for http/1.0 I'd like to keep for a while. I will drop it if I will not find a use case for it.

jogme · 2025-11-05T14:20:55Z

source/ssl_poll_perf.c

+        return NULL;
+
+    switch (alpn) {
+    case 0:


use the enum types here instead of magic numbers; other places also

source/ssl_poll_perf.c

jogme · 2025-11-05T17:17:36Z

source/ssl_poll_perf.c

+        break;
+    case RUN_MODE_CLIENT:
+        if (argv[optind] != NULL) {
+            url = argv[optind];


the parse_url can be done here and fail right away if the url is missing or incorrect

jogme · 2025-11-05T17:20:12Z

source/ssl_poll_perf.c

+    free(hostname);
+    free(portstr);
+    free(request_path);


these should be freed on the error path too

Co-authored-by: Pocs Norbert <norbertpocs0@gmail.com>

t8m · 2025-11-06T09:32:15Z

I'm wondering how to better run this in our performance benchmark lab. Is it enough to run both server and client on the same machine? Or the client and the server must be resided on separate hosts. In that case does it matter the platform of the counterpart? It will also bring questions of the simultaneous access to the counterpart from different testing machines. How will it affect on the results?

IMO it SHOULD run on the same machine over loopback to avoid network latencies, etc.

Sashan · 2025-11-06T09:45:58Z

I'm wondering how to better run this in our performance benchmark lab. Is it enough to run both server and client on the same machine? Or the client and the server must be resided on separate hosts. In that case does it matter the platform of the counterpart? It will also bring questions of the simultaneous access to the counterpart from different testing machines. How will it affect on the results?

this is should be sufficient:

lifty$ ./ssl_poll_perf -t /path/to/certs/servercert.pem /path/to/certs/serverkey.pem  
268.1150

The output is the time it took to establish 10 connections. each connection opens 10 bidirectional streams and 10 unidirectional stream pairs. each stream performs http/1.0 request to obtaub 64bytes. This is a default pre-set. We can either change defaults here:

3405     const char *ccountstr = "10";
3406     const char *bstreamstr = "10";
3407     const char *ustreamstr = "10";
3408     const char *rep_sizestr = "64";
3409     const char *req_sizestr = "64";

Or you can pass commandline options which control those settings. command below performs the same workload as ./ssl_poll_perf -t /path/to/certs//... comands

./ssl_poll_test -c 10 -b 10 -u 10 -w 64 -s 64 -t /path/to/certs/servercert.pem /path/to/certs/serverkey.pem

Or are you asking how to cross test OpenSSL quic stack with other QUIC clients/servers?

jogme · 2025-11-06T10:48:25Z

source/ssl_poll_perf.c

+         if (port_l != 0) {
+             free(portstr);
+             portstr = service;
+             service = NULL;


when no port is defined BIO_parse_hostserv fills service with u to me locally, strtol return 0 and the assert below on service fails. IMO moving this line outside the "if" should fix it

in this case assert() does not hold and we are better to put OPENSSL_free() back for service. I think it's more sensible.

okay, I have nothing against

jogme · 2025-11-06T10:50:51Z

source/ssl_poll_perf.c

+    case RUN_MODE_CLIENT:
+        if (argv[optind] != NULL)
+            parse_url(argv[optind]);
+        break;


this should overflow when the above if is false

In that case I prefer to restore what was there.

iirc that still does not correctly handle the case when there is no URL provided in the program args

Enable ssl_poll_perf test to run as client/server.

00a6865

This enables as to evaluate performance of OpenSSL QUIC stack against other QUIC implementations.

Sashan added this to Development Board Nov 5, 2025

Sashan moved this to Waiting Review in Development Board Nov 5, 2025

vavroch2010 assigned Sashan Nov 5, 2025

Sashan requested review from nhorman and quarckster and removed request for quarckster November 5, 2025 12:20

Sashan assigned nhorman and jogme Nov 5, 2025

jogme suggested changes Nov 5, 2025

View reviewed changes

Sashan and others added 4 commits November 6, 2025 09:52

Apply suggestion from @jogme

cc79b78

Co-authored-by: Pocs Norbert <norbertpocs0@gmail.com>

Update source/ssl_poll_perf.c

1708048

Co-authored-by: Pocs Norbert <norbertpocs0@gmail.com>

Update source/ssl_poll_perf.c

9d4992a

Co-authored-by: Pocs Norbert <norbertpocs0@gmail.com>

Update source/ssl_poll_perf.c

c7ba454

Co-authored-by: Pocs Norbert <norbertpocs0@gmail.com>

- fixing nits pointed out by Norbert

09b030c

Sashan requested a review from jogme November 6, 2025 10:06

jogme suggested changes Nov 6, 2025

View reviewed changes

- address issue pointed out by Norbert

f54a871

Sashan requested a review from jogme November 6, 2025 11:50

vavroch2010 requested review from mattcaswell and t8m November 11, 2025 09:03

Enable ssl_poll_perf test to run as client/server. #63

Are you sure you want to change the base?

Enable ssl_poll_perf test to run as client/server. #63

Conversation

Sashan commented Nov 5, 2025

Uh oh!

Sashan commented Nov 5, 2025

Uh oh!

quarckster commented Nov 5, 2025

Uh oh!

jogme left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

t8m commented Nov 6, 2025

Uh oh!

Sashan commented Nov 6, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants