And in my totally unscientific tests, it rocks!
Do we want to bump some of the other max settings in the process?
As mentioned by Kohei in bug#418962#c3 Fasterfox tweaks these settings:
Those settings would get us banned from the internet. Let's deal with other settings in other bugs, so that the known-wanted tweaks don't get held up behind research.
Let's not scope creep, here, and keep this bug about max-persistent-connections.
Firefox 2: 2
Opera 9.26: 4
Opera 9.5 beta: 4
Safari 3.0.4 Mac/Windows: 4
IE 7: 2
IE 8: 6
Based on that I'm wondering if we should be conservative and just double it to 4. Rob, does that still rock?
Blocking on resolution one way or the other. (Personally, I think we're safe doing whatever IE8 does, although it's hard to say if that's what they'll ship in their final version or not.)
-> sayrer (this is just a pref tweak AIUI)
They also provide a scriptable interface to give sites some more control, which we won't be aping. We can reduce the default in a .0.x patch if all hell breaks loose, but I think it rather unlikely; app developers seem pretty universally in favour of an increased limit.
Limiting this to just max-persistent-connections without raising network.http.max-connections or network.http.max-connections-per-server may not make much sense. If we're allowing 6 persistent connections but only 8 connections total, that seems odd to me. As I understand, IE upped both limits.
The current numbers in all.js are:
So it looks like Fasterfox triples the two smallest numbers, doubles the 8, and bumps up the 24 "some".
I would advocate doing at least:
I think we could even go to 6 on max-persistent-connections with max-connections at 12, though at that point 16 might make more sense.
(In reply to comment #8)
> pref("network.http.max-persistent-connections-per-server", 4);
> pref("network.http.max-persistent-connections-per-proxy", 8);
Why would you make these different?
Because if you're being proxied then _all_ your network requests are going to the proxy. If you browse multiple sites, tough. You're limited by the proxy cap, no questions asked.
(In reply to comment #10)
> Because if you're being proxied then _all_ your network requests are going to
> the proxy.
Not if you're using the FoxyProxy extension, but point taken.
Sure. Or you could be using a PAC, or whatever. But in many cases (e.g. corporate setups) the user is just being proxied, period.
An interesting article about Testing IE8’s Connection Parallelism:
1) When using connection parallelism tricks (spreading resources on
image1.example.com, image2.example.com, image2.example.com all pointing to your
server), the performance can be reduced by the large number of connections
opened (n_hosts * n_connections_per_host).
2) IE 8 is apparently going to handle limits differently for dial-up: "Kris Zyp
and Steve Souders have pointed out that IE8 will use 2 connections per host for
I think that authors should be taught that doing 1) could be harmful if they do
not take into account the default connection settings of the browsers they are
dealing with (yay for UA sniffing).
2) would be a good enhancement for dial-up users. But I guess they are a
minority these days.
(In reply to comment #7)
> They also provide a scriptable interface to give sites some more control, which
> we won't be aping.
As far as I could tell from looking at the whitepaper, it's a read-only property. So sites can use it to throttle their XHRs etc but they can't change it.
Created attachment 310171 [details] [diff] [review]
The thinking here is this: our limits are antiquated. Sites are demanding more connections due to JS traffic, and we're using up more of the total connection limit with suggest requests, antimalware, extension traffic, etc.
Comment on attachment 310171 [details] [diff] [review]
Alea iacta est. Or at least sorta.
Comment on attachment 310171 [details] [diff] [review]
Sure, let's try this for b5.
This didn't seem to affect Tp, right? Connection too fast and too low-latency there to notice?
Obviously we need some sort of Tgmail.
This change might induce problems with Apache servers. Each connection kept alive on an apache server takes a "processing unit" (be it a thread or a process), which means that taking 6 connections instead of 2 will multiply by 3 the number of processing units used on servers. Except for an experimental apache engine, there is no pooling of kept alive connections, so each kept alive connection prevents a processing unit to give content to other browsers.
And considering it is still recommended to use the prefork mpm (where processing units are processes) for php (because it is not thread safe), increasing the number of processes may not be an option.
Surely, anyone can easily DoS a web site running apache by opening a lot of keep-alive connections, but it would be better if genuine browsers wouldn't do it.
(In reply to comment #20)
> This change might induce problems with Apache servers.
Surely such problems would have already been observed with Safari 3 and Opera 9 (which both allow 4 persistent connections). Betas are the best place to try these changes.
> will multiply by 3 the number of processing units used on servers.
I'd love to hear your reasoning for that conclusion. Here's why I think it's not quite the right conclusion.
What will be multiplied by 3 is the number of simultaneous persistent connections a given user can hold open to the server. If we assume that this change doesn't affect the number of users, the average number of total connections held open on the server will be scaled by the product of two ratios: The ration of the average number of connections held open to the server after this change to the number held open before this change and the ratio of average time a connection is held open after this change to the average time a connection is held open before this change.
During pageload, either the loads will be parallelized over more connections (with pageload being faster as a result), or they won't be (and then you don't have more connections). In either case, I would expect the product to be close to unity for this scenario. The client will hold more connections open to the server, perhaps, but hold them open for less time. There is an obvious limiting factor here, which is network bandwidth: if the user can't actually benefit speed-wise from more connections, then the server loses. There are plans to dynamically lower the limit on dial-up connections.
Similar arguments would apply to pretty much every workload that I can realistically think of where time-averaging is applicable. The only exceptions I've been able to think of are:
1) A server which has all its users hitting it all at once (download after
some announcement, presumably, since page loads are too transient for this
to happen on any large scale).
2) A server to which users connect and then suck down arbitrary amounts of
data using as many connections as they can, with the connection lifetimes
fixed. I'm not sure why this would happen. The users would be getting 3
times the data they get now (again, assuming bandwidth is not the limiter),
so I'd expect them to be connecting 3 times less often, on average, in any
There's also the minor issue that every single other browser has the limit higher than 2, so Apache loses no matter what we do.
(In reply to comment #22)
> The client will hold more connections open to the
> server, perhaps, but hold them open for less time.
Keep-alive usually holds the connections much longer than download time. Holding 6 connections alive is going to keep 6 connections open for the keep-alive timeout (15 seconds by default on Apache), while it would have been 2 before. Or do you mean keep-alive is disabled ?
(In reply to comment #21)
> (In reply to comment #20)
> > This change might induce problems with Apache servers.
> Surely such problems would have already been observed with Safari 3 and Opera 9
> (which both allow 4 persistent connections). Betas are the best place to try
> these changes.
How much "market share" do Safari 3 and Opera 9 have ? How likely is it that there are enough Safari 3 and Opera 9 users on a same server that it would be a big difference ? Same applies with betas. You won't know how broken your choice was until Apache admins are forced to disable keep-alive on their servers.
No, keep-alive is enabled (otherwise we wouldn't be talking about persistent connections). So in the above calculations, the right ratio is
(lifetime + 15s)/(lifetime + 15s) instead of lifetime/lifetime. This does skew the results somewhat, though I'd be interested in what typical lifetimes look like. We're still not likely to be saturating the 6 connections on a lot of pages (due to blocking on scripts), and Apache still has this issue no matter what we do.
As for the rest, so far the plan is to ship this in a beta. Then we'll see.
Of course if there is hard data that this causes problems in the beta, we'd love to hear about it!
If you want to be safe, you'd better close connections when there are several waiting for keep-alive timeout on the same server. You're not really likely to require all of them during this time. You're usually not even using a single one. Anyways, one should be enough.
(better safe than sorry, because sorry would be too late)
I am surprised no one has mentioned section 8.1.4 of the HTTP 1.1 RFC:
quote: "A single-user client SHOULD NOT maintain more than 2 connections with any server or proxy."
(In reply to comment #29)
> I am surprised no one has mentioned section 8.1.4 of the HTTP 1.1 RFC:
I've seen the document once or twice. Which browsers comply with the section you've quoted?
(In reply to comment #30)
> I've seen the document once or twice. Which browsers comply with the section
> you've quoted?
I think mike already answered that in comment #4.
I do think it's worth documenting in this bug where the '2 connections' originally came from, and why it's been like that for so long.
Since the RFC says SHOULD and not MUST, that leaves some leeway to increase it.
(In reply to comment #26)
> Of course if there is hard data that this causes problems in the beta, we'd
> love to hear about it!
Are there any identified areas which are likely to see high FF3b5 user percentages? Can anyone reach out to them and see if they can report before/after performance? I'm thinking, for example, AMO, mozillazine.org, planet.m.o?
(In reply to comment #20)
> the number of processing units used on servers. Except for an experimental
> apache engine, there is no pooling of kept alive connections, so each kept
> alive connection prevents a processing unit to give content to other browsers.
> And considering it is still recommended to use the prefork mpm (where
> processing units are processes) for php (because it is not thread safe),
> increasing the number of processes may not be an option.
> Surely, anyone can easily DoS a web site running apache by opening a lot of
> keep-alive connections, but it would be better if genuine browsers wouldn't do
Running a reverse-proxy (Squid) or todays favourite's (nginx,Varnish, lighttpd 1.5) in front of Apache has always been best practice. Let Apache listen on 127.0.0.1, turn Keep-Alive Off and front it with nginx who does all the keepalive management
moving up from 3.0.4 to running the 3.1b2 candidate builds in the last week I've seen a lot of strange network behavior with attempts to visit sites timing out. then hit reload once or twice and the site loads ok. All very hard to preproduce in any way. I don't know if it is flakey connections at home or work, or what it might be. Is there any chance that upping the connections might create race conditions or logjams in network loading? I'll keep trying to see if the is a way to reproduce something, but we should watch for this kind of feedback from others, and if we do figure out what changes it might be connected to.
Chris, you could try setting these prefs to the old values locally and seeing whether that has an effect....
Also, FWIW, we fixed this in Firefox 3.0, so there should be no difference between that and 3.1b2.
re: last three comments.
I changed back to older connection levels and didn't see any change in behavior. Still lots of annoying time-outs. I also changed back to 3.0.4 and don't seem to see the problem there.
One specfic place I saw a high pct. of problems were with 3.1b2 was when I clicking on links to wikipedia pages from google search results pages. I also noticed that the delay might be early in the cycle of trying to connect to a site, and status bar messages that indicate it might be in the stages of DNS lookups. Maybe bugs 453403 and 464838 might be behind my problem.
(In reply to comment #37)
> One specfic place I saw a high pct. of problems were with 3.1b2 was when I
> clicking on links to wikipedia pages from google search results pages. I also
> noticed that the delay might be early in the cycle of trying to connect to a
> site, and status bar messages that indicate it might be in the stages of DNS
> lookups. Maybe bugs 453403 and 464838 might be behind my problem.
Are you maybe seeing bug 467562?
I'm seeing another problem here with this setting. I've got a small embedded device - a VoIP phone Siemens Gigaset S685 - which just gets flooded by FF 3.1b3 with requests. By setting network.http.max-connections-per-server to 15 or even to 8 the web ui of the device becomes unusable. Only setting it to 4 makes the web ui usable again.
If I understand the comments correctly bumping the number of concurrent connections per server should improve sites which use a lot of different JS scripts. But doesn't changing the browsers networking behavior "fix" this problem in the wrong place? Shouldn't the app developers combine and compress their JS scripts to a single file so that the browser doesn't have to load lots of JS files?
In any case setting this value so high breaks other applications to the point where they are unusable and not easy to fix.
Running FF 3.1b3 on a Mac.
> If I understand the comments correctly bumping the number of concurrent
> connections per server should improve sites which use a lot of different JS
I don't think you understand correctly. The goal is to improve any site which uses more than the previous max-connections-per-server number of resources on the page (images, stylesheets, etc). Note that the value was 8 for a very very long time, so if that value doesn't work with your device then it's not like this patch made things any worse. Note that all modern browsers use values that are at least this high.
It really sounds like your real problem is a severely buggy HTTP server on the device in question...