Repository navigation
HTTP/2 downloads slower with default flow control #1102
Description
Downloading a large object is much slower using HTTP/2. I peeked at the code and I'm pretty sure curl is using the default flow control window (65k). This is pretty small (I think Chrome opens the window to 10MB? by default). I didn't see an option to use a larger flow control window. I suggest that curl use a larger buffer by default and provide an option to use a larger one. Here's an example of pulling a 23MB image with a 33ms RTT:
HTTP/1.1
$ time curl https://upload.wikimedia.org/wikipedia/commons/d/d6/Seattle_7.jpg >/dev/null --http1.1
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
100 23.0M 100 23.0M 0 0 21.1M 0 0:00:01 0:00:01 --:--:-- 21.1M
real 0m1.101s
user 0m0.084s
sys 0m0.113s
HTTP/2
$ time curl https://upload.wikimedia.org/wikipedia/commons/d/d6/Seattle_7.jpg >/dev/null
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
100 23.0M 100 23.0M 0 0 2858k 0 0:00:08 0:00:08 --:--:-- 2945k
real 0m8.283s
user 0m0.111s
sys 0m0.097s
curl -V
curl 7.49.0 (x86_64-redhat-linux-gnu) libcurl/7.49.0 NSS/3.19.1 Basic ECC zlib/1.2.7 libidn/1.28 libpsl/0.7.0 (+libicu/50.1.2) libssh2/1.4.3 nghttp2/1.7.1
Protocols: dict file ftp ftps gopher http https imap imaps ldap ldaps pop3 pop3s rtsp scp sftp smb smbs smtp smtps telnet tftp
Features: AsynchDNS IDN IPv6 Largefile GSS-API Kerberos SPNEGO NTLM NTLM_WB SSL libz HTTP2 UnixSockets Metalink PSL
CentOS Linux release 7.2.1511 (Core)
Linux 4.0.9-68_fbk12_4058_gd84fdb3 #1 SMP Fri Aug 26 15:02:40 PDT 2016 x86_64 x86_64 x86_64 GNU/Linux
Activity
bagder commentedon Nov 3, 2016
Correct, we should start with fixing the h2 window use as mentioned in the TODO.
jay commentedon Nov 15, 2016
I can reproduce with that website but I tried retrieving the same file from a local server Apache/2.4.18 w/nghttp2 and the difference isn't significant it averaged .004s real faster for http1.1. Since it is local I guess that's to be expected. How do we do this, if I change
NGHTTP2_SETTINGS_INITIAL_WINDOW_SIZEto1 << 30I don't see any change in speed./cc @tatsuhiro-t
tatsuhiro-t commentedon Nov 16, 2016
The delay is caused by the latency between the moment sender blocked due to window size and the moment it receives WINDOW_UPDATE. The local server (loopback?) has low latency, so the difference is not significant.
I think you should also enlarge connection level flow control window by calling
nghttp2_session_set_local_window_size(session, NGHTTP2_FLAG_NONE, 0, 1 << 30)afternghttp2_submit_settings().jay commentedon Nov 16, 2016
Yup that did it, thanks. It only works if I do both. Here is the first draft https://github.com/curl/curl/compare/master...jay:use_large_http2_window?expand=1
bagder commentedon Nov 16, 2016
Looks good to me!
tatsuhiro-t commentedon Nov 16, 2016
Looks good to me too.
http2: Use huge HTTP/2 windows
jay commentedon Nov 16, 2016
Thanks guys. I just landed this in a4d8888 from a second draft, which has these changes from draft1:
HTTP2_HUGE_WINDOW_SIZEfor the huge window size so if it needs to be changed it will affect both window size settings. In draft1 I just used 1 << 30 two places.Curl_http2_switched. In draft1 I was not setting it for upgraded connections.bagder commentedon Nov 16, 2016
Great work @jay, I've wanted to see this done for a while but I never manged to pull myself together to get it done. Now you did it. Thanks!
afrind commentedon Nov 16, 2016
Thanks all!
From: Daniel Stenberg notifications@github.com
Reply-To: curl/curl reply@reply.github.com
Date: Wednesday, November 16, 2016 at 2:49 PM
To: curl/curl curl@noreply.github.com
Cc: Alan Frindell afrind@fb.com, Author author@noreply.github.com
Subject: Re: [curl/curl] HTTP/2 downloads slower with default flow control (#1102)
Great work @jayhttps://github.com/jay, I've wanted to see this done for a while but I never manged to pull myself together to get it done. Now you did it. Thanks!
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHubhttps://github.com//issues/1102#issuecomment-261098180, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AH4IabHpjLD2vFXT00Ii4ol46TtOWd9sks5q-4iKgaJpZM4KonU9.