Skip parsing after repeated partials on the same token
When the parse buffer contains the starting bytes of a token but not
all of them, we cannot parse the token to completion. We call this a
partial token. When this happens, the parse position is reset to the
start of the token, and the parse() call returns. The client is then
expected to provide more data and call parse() again.
In extreme cases, this means that the bytes of a token may be parsed
many times: once for every buffer refill required before the full token
is present in the buffer.
Math:
Assume there's a token of T bytes
Assume the client fills the buffer in chunks of X bytes
We'll try to parse X, 2X, 3X, 4X ... until mX == T (technically >=)
That's (m²+m)X/2 = (T²/X+T)/2 bytes parsed (arithmetic progression)
While it is alleviated by larger refills, this amounts to O(T²)
Expat grows its internal buffer by doubling it when necessary, but has
no way to inform the client about how much space is available. Instead,
we add a heuristic that skips parsing when we've repeatedly stopped on
an incomplete token. Specifically:
* Only try to parse if we have a certain amount of data buffered
* Every time we stop on an incomplete token, double the threshold
* As soon as any token completes, the threshold is reset
This means that when we get stuck on an incomplete token, the threshold
grows exponentially, effectively making the client perform larger buffer
fills, limiting how many times we can end up re-parsing the same bytes.
Math:
Assume there's a token of T bytes
Assume the client fills the buffer in chunks of X bytes
We'll try to parse X, 2X, 4X, 8X ... until (2^k)X == T (or larger)
That's (2^(k+1)-1)X bytes parsed -- e.g. 15X if T = 8X
This is equal to 2T-X, which amounts to O(T)
We could've chosen a faster growth rate, e.g. 4 or 8. Those seem to
increase performance further, at the cost of further increasing the
risk of growing the buffer more than necessary. This can easily be
adjusted in the future, if desired.
This is all completely transparent to the client, except for:
1. possible delay of some callbacks (when our heuristic overshoots)
2. apps that never do isFinal=XML_TRUE could miss data at the end
For the affected testdata, this change shows a 100-400x speedup.
The recset.xml benchmark shows no clear change either way.
Before:
benchmark -n ../testdata/largefiles/recset.xml 65535 3
3 loops, with buffer size 65535. Average time per loop: 0.270223
benchmark -n ../testdata/largefiles/aaaaaa_attr.xml 4096 3
3 loops, with buffer size 4096. Average time per loop: 15.033048
benchmark -n ../testdata/largefiles/aaaaaa_cdata.xml 4096 3
3 loops, with buffer size 4096. Average time per loop: 0.018027
benchmark -n ../testdata/largefiles/aaaaaa_comment.xml 4096 3
3 loops, with buffer size 4096. Average time per loop: 11.775362
benchmark -n ../testdata/largefiles/aaaaaa_tag.xml 4096 3
3 loops, with buffer size 4096. Average time per loop: 11.711414
benchmark -n ../testdata/largefiles/aaaaaa_text.xml 4096 3
3 loops, with buffer size 4096. Average time per loop: 0.019362
After:
./run.sh benchmark -n ../testdata/largefiles/recset.xml 65535 3
3 loops, with buffer size 65535. Average time per loop: 0.269030
./run.sh benchmark -n ../testdata/largefiles/aaaaaa_attr.xml 4096 3
3 loops, with buffer size 4096. Average time per loop: 0.044794
./run.sh benchmark -n ../testdata/largefiles/aaaaaa_cdata.xml 4096 3
3 loops, with buffer size 4096. Average time per loop: 0.016377
./run.sh benchmark -n ../testdata/largefiles/aaaaaa_comment.xml 4096 3
3 loops, with buffer size 4096. Average time per loop: 0.027022
./run.sh benchmark -n ../testdata/largefiles/aaaaaa_tag.xml 4096 3
3 loops, with buffer size 4096. Average time per loop: 0.099360
./run.sh benchmark -n ../testdata/largefiles/aaaaaa_text.xml 4096 3
3 loops, with buffer size 4096. Average time per loop: 0.017956
2 files changed
