Re: [racket-users] Re: note about parsing speed of xml vs sxml?
I tested the your string port version and I also wrote a "string-append" version of the xml reader and they are both slower by about 10-15% on my machine, when compared to the current read-xml implementation which uses `list->string`. It looks like `list->string` is not the bottleneck here. There are some small improvements that can be made from micro optimizations. For example, I changed `name-char?` to not use `name-start?` but instead check for all chars, and I also changed `lex-name` to construct the list in reverse and use `(list->string (reverse chars))`, plus I reordered the cond condition to check the common case first (that the next character is a name-char? and not a 'special one). However, this resulted in only about 5-10% speed improvement, nowhere near where the 4 time speedup when using `sxml`, as reported by John. In the end, it may well be that speeding up `read-xml` can only be done by these types of micro-optimizations. Another thing I looked into was the "pattern" used for reading: all the `read-xml` code will use the pattern of "peeking" the next character, deciding if it is good, than reading it. This is much slower than just reading the characters directly. These are the results from just reading in a 14Mb XML file: read-char only: cpu time: 312 real time: 307 gc time: 0 read-char-or-special only: cpu time: 750 real time: 741 gc time: 0 peek-char than read-char: cpu time: 1234 real time: 1210 gc time: 0 peek-char-or-special than read-char-or-special: cpu time: 1688 real time: 1690 gc time: 0 Using this code: (define file-name "your-test-file-here.xml") (printf "read-char only~%") (collect-garbage 'major) (time (call-with-input-file file-name (lambda (in) (let loop ([c (read-char in)]) (if (eof-object? c) (void) (loop (read-char in))) (printf "read-char-or-special only~%") (collect-garbage 'major) (time (call-with-input-file file-name (lambda (in) (let loop ([c (read-char-or-special in)]) (if (eof-object? c) (void) (loop (read-char-or-special in))) (printf "peek-char than read-char~%") (collect-garbage 'major) (time (call-with-input-file file-name (lambda (in) (let loop ([c (peek-char in)]) (if (eof-object? c) (void) (begin (void (read-char in)) (loop (peek-char in (printf "peek-char-or-special than read-char-or-special~%") (collect-garbage 'major) (time (call-with-input-file file-name (lambda (in) (let loop ([c (peek-char-or-special in)]) (if (eof-object? c) (void) (begin (void (read-char-or-special in)) (loop (peek-char-or-special in Alex. On Monday, June 29, 2020 at 5:30:43 AM UTC+8 rmculp...@gmail.com wrote: > Thanks Alex for pointing out the use of list->string. I've created a PR ( > https://github.com/racket/racket/pull/3275) that changes that code to use > string ports instead (similar to Hendrik's suggestion, but the string port > handles resizing automatically). Could someone (John?) with some large XML > files lying around try the changes and see if they help? > > Ryan > > > On Sun, Jun 28, 2020 at 9:56 PM Neil Van Dyke > wrote: > >> If anyone wants to optimize `read-xml` for particular classes of use, >> without changing the interface, it might be very helpful to run your >> representative tests using the statistical profiler. >> >> The profiler text report takes a little while of tracing through >> manually to get a feel for how to read and use it, but it can be >> tremendously useful, and is worth learning to do if you need performance. >> >> After a first pass with that, you might also want to look at how costly >> allocations/GC are, and maybe do some controlled experiments around >> that. For example, force a few GC cycles, run your workload under >> profiler, check GC time during, and forced time after. If you're >> dealing with very large graphs coming out of the parser, I don't know >> whether those are enough to matter with the current GC mechanism, but >> maybe also check GC time while you're holding onto large graphs, when >> you release them, and after they've been collected. At some point, GC >> gets hard for at least me to reason about, but some things make sense, >> and other things you decide when to stop digging. :) If you record all >> your measurements, you can compare empirically the how different changes >> to the code affect things, hopefully in representative situations. >> >> I went through a lot of these exercises to optimize a large system, and >> sped up dynamic Web page loads dramatically in the usual case (to the >> point we were then mainly limited by PostgreSQL query cost, not much by >> the application code in Scheme, nor our request&response network I/O), >> and also greatly reduced the pain of intermittent request latency spikes >> due to GC. >> >> O
Re: [racket-users] Re: note about parsing speed of xml vs sxml?
I suggested using `string-append` because in my own performance investigations with reading 100Mb+ CSV files: constructing short tokens using string-append is faster than using a string port -- perhaps there is a fixed overhead with using string ports which makes `string-append` faster for short strings, but I don't know at what string length the string ports become faster. I think using string ports will be definitely faster than using `list->string`, but for the difference between `string-append` and string ports, some performance measurement might be needed. Thanks for looking into this, Alex. On Monday, June 29, 2020 at 5:30:43 AM UTC+8 rmculp...@gmail.com wrote: > Thanks Alex for pointing out the use of list->string. I've created a PR ( > https://github.com/racket/racket/pull/3275) that changes that code to use > string ports instead (similar to Hendrik's suggestion, but the string port > handles resizing automatically). Could someone (John?) with some large XML > files lying around try the changes and see if they help? > > Ryan > > > On Sun, Jun 28, 2020 at 9:56 PM Neil Van Dyke > wrote: > >> If anyone wants to optimize `read-xml` for particular classes of use, >> without changing the interface, it might be very helpful to run your >> representative tests using the statistical profiler. >> >> The profiler text report takes a little while of tracing through >> manually to get a feel for how to read and use it, but it can be >> tremendously useful, and is worth learning to do if you need performance. >> >> After a first pass with that, you might also want to look at how costly >> allocations/GC are, and maybe do some controlled experiments around >> that. For example, force a few GC cycles, run your workload under >> profiler, check GC time during, and forced time after. If you're >> dealing with very large graphs coming out of the parser, I don't know >> whether those are enough to matter with the current GC mechanism, but >> maybe also check GC time while you're holding onto large graphs, when >> you release them, and after they've been collected. At some point, GC >> gets hard for at least me to reason about, but some things make sense, >> and other things you decide when to stop digging. :) If you record all >> your measurements, you can compare empirically the how different changes >> to the code affect things, hopefully in representative situations. >> >> I went through a lot of these exercises to optimize a large system, and >> sped up dynamic Web page loads dramatically in the usual case (to the >> point we were then mainly limited by PostgreSQL query cost, not much by >> the application code in Scheme, nor our request&response network I/O), >> and also greatly reduced the pain of intermittent request latency spikes >> due to GC. >> >> One of the hotspots, I did half a dozen very different implementations, >> including C extension, and found an old-school pure Scheme >> implementation was fastest. I compared the performance of the >> implementation using something like `shootout`, but there might be >> better ways now in Racket. https://www.neilvandyke.org/racket/shootout/ >> I also found we could be much faster if we made a change to what the >> algorithm guarantees, since it was more of a consistency check that >> turned out to be very expensive and very redundant, due to all the ways >> that utility code ended up being used. >> >> In addition to contrived experiments, I also rigged up a runtime option >> so that the server would save data from the statistical profiler for >> each request a Web server handled in production. Which was tremendously >> useful, since it gave us real-world examples that were also difficult to >> synthesize (e.g., complex dynamic queries), and we could go from Web >> logs and user feedback, to exactly what happened. >> >> (In that system I optimized, we used Oleg's SXML tools very heavily >> throughout the system, plus some bespoke SXML tools for HTML and XML. >> There was one case in which someone had accidentally used the `xml` >> module, not knowing it was incompatible with the rest of the system, >> which caused some strange failures (no static checking) before it was >> discovered, and we changed that code to use SXML.) >> >> -- >> You received this message because you are subscribed to the Google Groups >> "Racket Users" group. >> To unsubscribe from this group and stop receiving emails from it, send an >> email to racket-users...@googlegroups.com. >> To view this discussion on the web visit >> https://groups.google.com/d/msgid/racket-users/68624c9a-df35-14a3-a912-df806799a7e0%40neilvandyke.org >> . >> > -- You received this message because you are subscribed to the Google Groups "Racket Users" group. To unsubscribe from this group and stop receiving emails from it, send an email to racket-users+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.co
Re: [racket-users] Re: note about parsing speed of xml vs sxml?
On Sun, Jun 28, 2020 at 11:30:27PM +0200, Ryan Culpepper wrote: > Thanks Alex for pointing out the use of list->string. I've created a PR ( > https://github.com/racket/racket/pull/3275) that changes that code to use > string ports instead (similar to Hendrik's suggestion, but the string port > handles resizing automatically). Could someone (John?) with some large XML > files lying around try the changes and see if they help? I'm currently using sxml to read the current openGL specification. 2564178 bytes in one file. (a) Is that relevant to your request? (b) What do I do to have a choice of which sxml to use? (c) Do I need to figure out using xml as well? (d) And what do you want the test to do? -- hendrik > > Ryan > > > On Sun, Jun 28, 2020 at 9:56 PM Neil Van Dyke wrote: > > > If anyone wants to optimize `read-xml` for particular classes of use, > > without changing the interface, it might be very helpful to run your > > representative tests using the statistical profiler. > > > > The profiler text report takes a little while of tracing through > > manually to get a feel for how to read and use it, but it can be > > tremendously useful, and is worth learning to do if you need performance. > > > > After a first pass with that, you might also want to look at how costly > > allocations/GC are, and maybe do some controlled experiments around > > that. For example, force a few GC cycles, run your workload under > > profiler, check GC time during, and forced time after. If you're > > dealing with very large graphs coming out of the parser, I don't know > > whether those are enough to matter with the current GC mechanism, but > > maybe also check GC time while you're holding onto large graphs, when > > you release them, and after they've been collected. At some point, GC > > gets hard for at least me to reason about, but some things make sense, > > and other things you decide when to stop digging. :) If you record all > > your measurements, you can compare empirically the how different changes > > to the code affect things, hopefully in representative situations. > > > > I went through a lot of these exercises to optimize a large system, and > > sped up dynamic Web page loads dramatically in the usual case (to the > > point we were then mainly limited by PostgreSQL query cost, not much by > > the application code in Scheme, nor our request&response network I/O), > > and also greatly reduced the pain of intermittent request latency spikes > > due to GC. > > > > One of the hotspots, I did half a dozen very different implementations, > > including C extension, and found an old-school pure Scheme > > implementation was fastest. I compared the performance of the > > implementation using something like `shootout`, but there might be > > better ways now in Racket. https://www.neilvandyke.org/racket/shootout/ > > I also found we could be much faster if we made a change to what the > > algorithm guarantees, since it was more of a consistency check that > > turned out to be very expensive and very redundant, due to all the ways > > that utility code ended up being used. > > > > In addition to contrived experiments, I also rigged up a runtime option > > so that the server would save data from the statistical profiler for > > each request a Web server handled in production. Which was tremendously > > useful, since it gave us real-world examples that were also difficult to > > synthesize (e.g., complex dynamic queries), and we could go from Web > > logs and user feedback, to exactly what happened. > > > > (In that system I optimized, we used Oleg's SXML tools very heavily > > throughout the system, plus some bespoke SXML tools for HTML and XML. > > There was one case in which someone had accidentally used the `xml` > > module, not knowing it was incompatible with the rest of the system, > > which caused some strange failures (no static checking) before it was > > discovered, and we changed that code to use SXML.) > > > > -- > > You received this message because you are subscribed to the Google Groups > > "Racket Users" group. > > To unsubscribe from this group and stop receiving emails from it, send an > > email to racket-users+unsubscr...@googlegroups.com. > > To view this discussion on the web visit > > https://groups.google.com/d/msgid/racket-users/68624c9a-df35-14a3-a912-df806799a7e0%40neilvandyke.org > > . > > > > -- > You received this message because you are subscribed to the Google Groups > "Racket Users" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to racket-users+unsubscr...@googlegroups.com. > To view this discussion on the web visit > https://groups.google.com/d/msgid/racket-users/CANy33q%3DpZw9EPmZG%2Bdz5cRYMSP17Ofntq9JwFqVVoN8ZhO6POg%40mail.gmail.com. -- You received this message because you are subscribed to the Google Groups "Racket Users" group. To unsubscribe from this group and stop receiving emails from it, send an email to racket-use
Re: [racket-users] Re: note about parsing speed of xml vs sxml?
Thanks Alex for pointing out the use of list->string. I've created a PR ( https://github.com/racket/racket/pull/3275) that changes that code to use string ports instead (similar to Hendrik's suggestion, but the string port handles resizing automatically). Could someone (John?) with some large XML files lying around try the changes and see if they help? Ryan On Sun, Jun 28, 2020 at 9:56 PM Neil Van Dyke wrote: > If anyone wants to optimize `read-xml` for particular classes of use, > without changing the interface, it might be very helpful to run your > representative tests using the statistical profiler. > > The profiler text report takes a little while of tracing through > manually to get a feel for how to read and use it, but it can be > tremendously useful, and is worth learning to do if you need performance. > > After a first pass with that, you might also want to look at how costly > allocations/GC are, and maybe do some controlled experiments around > that. For example, force a few GC cycles, run your workload under > profiler, check GC time during, and forced time after. If you're > dealing with very large graphs coming out of the parser, I don't know > whether those are enough to matter with the current GC mechanism, but > maybe also check GC time while you're holding onto large graphs, when > you release them, and after they've been collected. At some point, GC > gets hard for at least me to reason about, but some things make sense, > and other things you decide when to stop digging. :) If you record all > your measurements, you can compare empirically the how different changes > to the code affect things, hopefully in representative situations. > > I went through a lot of these exercises to optimize a large system, and > sped up dynamic Web page loads dramatically in the usual case (to the > point we were then mainly limited by PostgreSQL query cost, not much by > the application code in Scheme, nor our request&response network I/O), > and also greatly reduced the pain of intermittent request latency spikes > due to GC. > > One of the hotspots, I did half a dozen very different implementations, > including C extension, and found an old-school pure Scheme > implementation was fastest. I compared the performance of the > implementation using something like `shootout`, but there might be > better ways now in Racket. https://www.neilvandyke.org/racket/shootout/ > I also found we could be much faster if we made a change to what the > algorithm guarantees, since it was more of a consistency check that > turned out to be very expensive and very redundant, due to all the ways > that utility code ended up being used. > > In addition to contrived experiments, I also rigged up a runtime option > so that the server would save data from the statistical profiler for > each request a Web server handled in production. Which was tremendously > useful, since it gave us real-world examples that were also difficult to > synthesize (e.g., complex dynamic queries), and we could go from Web > logs and user feedback, to exactly what happened. > > (In that system I optimized, we used Oleg's SXML tools very heavily > throughout the system, plus some bespoke SXML tools for HTML and XML. > There was one case in which someone had accidentally used the `xml` > module, not knowing it was incompatible with the rest of the system, > which caused some strange failures (no static checking) before it was > discovered, and we changed that code to use SXML.) > > -- > You received this message because you are subscribed to the Google Groups > "Racket Users" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to racket-users+unsubscr...@googlegroups.com. > To view this discussion on the web visit > https://groups.google.com/d/msgid/racket-users/68624c9a-df35-14a3-a912-df806799a7e0%40neilvandyke.org > . > -- You received this message because you are subscribed to the Google Groups "Racket Users" group. To unsubscribe from this group and stop receiving emails from it, send an email to racket-users+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/racket-users/CANy33q%3DpZw9EPmZG%2Bdz5cRYMSP17Ofntq9JwFqVVoN8ZhO6POg%40mail.gmail.com.
Re: [racket-users] Re: note about parsing speed of xml vs sxml?
If anyone wants to optimize `read-xml` for particular classes of use, without changing the interface, it might be very helpful to run your representative tests using the statistical profiler. The profiler text report takes a little while of tracing through manually to get a feel for how to read and use it, but it can be tremendously useful, and is worth learning to do if you need performance. After a first pass with that, you might also want to look at how costly allocations/GC are, and maybe do some controlled experiments around that. For example, force a few GC cycles, run your workload under profiler, check GC time during, and forced time after. If you're dealing with very large graphs coming out of the parser, I don't know whether those are enough to matter with the current GC mechanism, but maybe also check GC time while you're holding onto large graphs, when you release them, and after they've been collected. At some point, GC gets hard for at least me to reason about, but some things make sense, and other things you decide when to stop digging. :) If you record all your measurements, you can compare empirically the how different changes to the code affect things, hopefully in representative situations. I went through a lot of these exercises to optimize a large system, and sped up dynamic Web page loads dramatically in the usual case (to the point we were then mainly limited by PostgreSQL query cost, not much by the application code in Scheme, nor our request&response network I/O), and also greatly reduced the pain of intermittent request latency spikes due to GC. One of the hotspots, I did half a dozen very different implementations, including C extension, and found an old-school pure Scheme implementation was fastest. I compared the performance of the implementation using something like `shootout`, but there might be better ways now in Racket. https://www.neilvandyke.org/racket/shootout/ I also found we could be much faster if we made a change to what the algorithm guarantees, since it was more of a consistency check that turned out to be very expensive and very redundant, due to all the ways that utility code ended up being used. In addition to contrived experiments, I also rigged up a runtime option so that the server would save data from the statistical profiler for each request a Web server handled in production. Which was tremendously useful, since it gave us real-world examples that were also difficult to synthesize (e.g., complex dynamic queries), and we could go from Web logs and user feedback, to exactly what happened. (In that system I optimized, we used Oleg's SXML tools very heavily throughout the system, plus some bespoke SXML tools for HTML and XML. There was one case in which someone had accidentally used the `xml` module, not knowing it was incompatible with the rest of the system, which caused some strange failures (no static checking) before it was discovered, and we changed that code to use SXML.) -- You received this message because you are subscribed to the Google Groups "Racket Users" group. To unsubscribe from this group and stop receiving emails from it, send an email to racket-users+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/racket-users/68624c9a-df35-14a3-a912-df806799a7e0%40neilvandyke.org.
Re: [racket-users] Re: note about parsing speed of xml vs sxml?
On Sat, Jun 27, 2020 at 05:16:34PM -0700, Alex Harsanyi wrote: > Looking at the source for `read-xml`, it seems to be using `list->string` > in several places. That is, it reads characters one-by-one and constructs > a list by appending a character to the end of it, than calls `list->string` > to produce the string. I suspect read-xml could be made faster by using > `string-append` in these cases. So you would be copying and reallocating strings instead of cons-cells? The way to make that eliminate all that allocation is to implement a likely big enough mutable string buffer and insert characters (likely one at at time if I read you correctly) without allocating new storage each time (unless you've made the buffer too smal; in which case, double its size). Then allocate the right amount of space for a string once and copy the buffer into it when the string has been completely read in. -- hendrik -- You received this message because you are subscribed to the Google Groups "Racket Users" group. To unsubscribe from this group and stop receiving emails from it, send an email to racket-users+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/racket-users/20200628190739.tpjrkilrf5eq546w%40topoi.pooq.com.
