On Wed, Oct 29, 2025 at 09:27:35AM +0100, Stefan Eissing wrote:
S> > On Tue, Oct 28, 2025 at 11:06:05PM +0100, Daniel Stenberg wrote:
S> > D> > This may actually help our case!
S> > D> > I'm going to give a try to unmodified 8.16.0.
...
S> > The regression was discovered by a huge A/B test of a system overall,
S> > that involves many machines, many hours and a ton of traffic. What
S> > I'm going to do now is to re-run this A/B test with 8.16.0, to avoid
S> > relying on year old data.
S>
S> Waiting for your results then. The "limit-rate revisisted" PR is
https://github.com/curl/curl/pull/18454
S>
S> I noticed "busy" loops where we expired transfers that are on a speed
limits. Hope that addresses the cpu issue for you. As to the accuracy, we'll
see what you find in your setup.
We had run our large test and I can confirm that 8.16.0 doesn't need our
patch any more!
I guess the chunk to transfer.c in 24badd29f is the key:
--- a/lib/transfer.c
+++ b/lib/transfer.c
@@ -280,9 +280,10 @@ static CURLcode sendrecv_dl(struct Curl_easy *data,
if(bytestoread && data->set.max_recv_speed > 0) {
/* In case of speed limit on receiving: if this loop already got
- * data, break out. If not, limit the amount of bytes to receive.
- * The overall, timed, speed limiting is done in multi.c */
- if(total_received)
+ * a quarter of the quota, break out. We want to stutter a bit
+ * to keep in the limit, but too small receives will just cost
+ * cpu unnecessarily. */
+ if(total_received >= (data->set.max_recv_speed / 4))
break;
if(data->set.max_recv_speed < (curl_off_t)bytestoread)
bytestoread = (size_t)data->set.max_recv_speed;
Thanks!
--
Gleb Smirnoff
--
Unsubscribe: https://lists.haxx.se/mailman/listinfo/curl-library
Etiquette: https://curl.se/mail/etiquette.html
