On Thu, Jan 7, 2010 at 12:49 PM, Maciej Stachowiak <m...@apple.com> wrote:
> > On Jan 7, 2010, at 12:09 PM, Mike Belshe wrote: > > Hi - > > I've been working on SPDY, but I think I may have found a good performance > win for HTTP. Specifically, if the PreloadScanner, which is responsible for > scanning ahead within an HTML document to find subresources, is throttled > today. The throttling is intentional and probably sometimes necessary. > Nonetheless, un-throttling it may lead to a 5-10% performance boost in some > configurations. I believe Antti is no longer working on this? Is there > anyone else working in this area that might have data on how aggressive the > PreloadScanner should be? Below I'll describe some of my tests. > > The PreloadScanner throttling happens in a couple of ways. First, the > PreloadScanner only runs when we're blocked on JavaScript (see > HTMLTokenizer.cpp). But further, as it discovers resources to be fetched, > it may delay or reject loading the subresource at all due to throttling in > loader.cpp and DocLoader.cpp. The throttling is very important, depending > on the implementation of the HTTP networking stack, because throwing too > many resources (or the low-priority ones) into the network stack could > adversely affect HTTP load performance. This latter problem does not impact > my Chromium tests, because the Chromium network stack does its own > prioritization and throttling (not too dissimilar from the work done by > loader.cpp). > > > The reason we do this is to prevent head-of-line blocking by low-priority > resources inside the network stack (mainly considering how CFNetwork / > NSURLConnection works). > Right - understood. > > > *Theory*: > The theory I'm working under is that when the RTT of the network is > sufficiently high, the *best* thing the browser can do is to discover > resources as quickly as possible and pass them to the network layer so that > we can get started with fetching. This is not speculative - these are > resources which will be required to render the full page. The SPDY > protocol is designed around this concept - allowing the browser to schedule > all resources it needs to the network (rather than being throttled by > connection limits). However, even with SPDY enabled, WebKit itself prevents > resource requests from fully flowing to the network layer in 3 ways: > a) loader.cpp orders requests and defers requests based on the state of > the page load and a number of criteria. > b) HTMLTokenizer.cpp only looks for resources further in the body when > we're blocked on JS > c) "preload" requests are treated specially (docloader.cpp); if they are > discovered too early by the tokenizer, then they are either queued or > discarded. > > > I think your theory is correct when SPDY is enabled, and possibly when > using HTTP with pipelining. It may be true to a lesser extent with > non-pipelining HTTP implementations when the network stack does its own > prioritization and throttling, by reducing latency in getting the request to > the network stack. > right. > This is especially so when issuing a network request to the network stack > may involve significant latency due to IPC or cross-thread communication or > the like. > I hadn't considered IPC or cross thread latencies. When I've measured these in the past they are very very low. One problem with the single-threaded nature of our preloader and parser right now is that if the HTMLTokenizer is in the middle of executing JS code, we're not doing anything to scan for preloads; tons of data can be flowing in off the network which we're oblivious to. I'm not trying to change this for now, though, it's much more involved, I think, due to thread safety requirements for the webcore cache. > > > *Test Case* > Can aggressive preloadscanning (e.g. always preload scan before parsing an > HTML Document) improve page load time? > > To test this, I'm calling the PreloadScanner basically as the first part of > HTMLTokenizer::write(). I've then removed all throttling from loader.cpp > and DocLoader.cpp. I've also instrumented the PreloadScanner to measure its > effectiveness. > > *Benchmark Setup* > Windows client (chromium). > Simulated network with 4Mbps download, 1Mbps upload, 100ms RTT, 0% packet > loss. > I run through a set of 25 URLs, loading each 30 times; not recycling any > connections and clearing the cache between each page. > These are running over HTTP; there is no SPDY involved here. > > > I'm interested in the following: > > - What kind of results do you get in Safari? > I've not done much benchmarking in Safari; do you have a good way to do this? Is there something I can read about or tools I can use? For chromium, I use the benchmarking extension which lets me run through lots of pages quickly. > - How much of this effect is due to more aggressive preload scanning and > how much is due to disabling throttling? Since the test includes multiple > logically indpendent changes, it is hard to tell which are the ones that had > an effect. > Great question. The chromium throttler is similar to the WebCore throttler - same number of connections, and same relative priority of subresources. So I just tested; if I only turn off throttling, the performance of chromium is unchanged from the baseline (only off by 3ms!). That is good - it means the chromium throttling is effectively the same as what webkit is providing. But for HTTP; the key seems to the pre-rendering-ready escape hatch in DocLoader::preload. Removing this gives me most all of the benefit. The comment says it pretty clearly: "Don't preload images or body resources before we have something to draw. This prevents preloads from body delaying first display when bandwidth is limited." For SPDY, there is more benefit by continuing to preparse aggressively - I suspect this is due to the finer grained prioritization where it can continue to send requests up without impacting the clogged downlink channel. So I ran the test on a 400Kbps link (ouch, that is painful). Overall PLT improved a marginally (this surprised me). But time-to-first-paint dropped by ~20%. More tests coming. > > *Results:* > Baseline > (without my changes)UnthrottledNotesAverage PLT2377ms2239ms+5.8% latency > redux.Time spent in the PreloadScanner 1160ms4540msAs expected, we spend > about 4x more time in the PreloadScanner. In this test, we loaded 750 pages, > so it is about 6ms per page. My machine is fast, though.Preload Scripts > discovered 262194404x more scripts discoveredPreload CSS discovered34810223x > more CSS discoveredPreload Images discovered1195239144 3x more images > discoveredPreload items throttled99830Preload Complete hits38036950This is > the count of items which were completely preloaded before WebKit even tried > to look them up in the cache. This is pure goodness. Preload Partial hits > 17087230These are partial hits, where the item had already started > loading, but not finished, before WebKit tried to look them up.Preload > Unreferenced 42130These are bad and the count should be zero. I'll try to > find them and see if there isn't a fix - the PreloadScanner is just > sometimes finding resources that are never used. It is likely due to clever > JS which changes the DOM. > > > > *Conclusions:* > For this network speed/client processor, more aggressive PreloadScanning > clearly is a win. More testing is needed for slower machines and other > network types. I've tested many network types; the aggressive preload > scanning seems to always be either a win or a wash; for very slow network > connections, where we're already at capacity, the extra CPU burning is > basically free. For super fast networks, with very low RTT, it also appears > to be a wash. The networks in the middle (including mobile simulations) see > nice gains. > > *Next Steps and Questions:* > I'd like to land my changes so that we can continue to gather data. I can > enable these via macro definitions or I can enable these via dynamic > settings. I can then try to do more A/B testing. > > > I'd like answers to my questions above before we consider that. > > > Are there any existing web pages which the WebKit team would like tested > under these configurations? I don't see a lot of testing that I can > leverage from the initial great work Antti did for verifying that I'm not > breaking anything. > > Is there any other information or data from the original PreloadScanner > work which I should read? > > > There's the original blog announcement of preload scanning: > > http://webkit.org/blog/166/optimizing-page-loading-in-web-browser/ > > It might be a good idea to try replicating those results with proposed > changes. > Thanks for the link - I had forgotten about that post... I'll add more tests for sure. This work is dependent both on the page type as well as the network properties. Testing only the wsj is testing only one type of html/css/js layout. I'm not even sure if the wsj basic structure has changed from when Antti did his tests... The tests I'm doing hits 25 major websites and doesn't use the internet. I can add wsj to my mix. Mike > Regards, > Maciej > >
_______________________________________________ webkit-dev mailing list webkit-dev@lists.webkit.org http://lists.webkit.org/mailman/listinfo.cgi/webkit-dev
