[22 August 2008]

My brief hope (it would be misleading to refer to it as a “plan”) to report daily from Balisage 2008 has bitten the dust — it did that shortly after noon on the first day, when my account of Sandro Hawke’s work on XTAN turned out to take more time than I had available — but there is still a lot to say. I’m going to abandon the chronological approach, however, and write about things that come to mind, in a more or less random order.

One of my favorite papers this year was the one submitted by Yu Wu, Qi Zhang, Zhiqiang Yu, and Jianhui Li, of Intel, under the slightly daunting title “A Hybrid Parallel Processing for XML Parsing and Schema Validation”. (I think they are all members of the XML Engineering Team at the Intel China Software Center in Shanghai, but I am not sure I’ve read all the affiliation info correctly; I keep being distracted by the implications of an Intel software center having an XML engineering team.)

When I paraphrased this paper to a friend recently, her response was “Wow! That’s a lot more accessible than I would have guessed from the title.” So perhaps it’s worth while to try to record here the high points of the work, in a way that’s accessible to people to people with no more than lay knowledge of the relevant technical disciplines. (This goal is made easier, of course, by the fact that I don’t have more than lay knowledge of those disciplines myself.)

For technical details, of course, readers should go to the paper in the online proceedings of the conference; all errors in this summary are mine.

The elevator speech

The quick executive summary: XML parsing, and validation, can be a lot faster if performed by a multi-threaded process using optimistic concurrency.

By “optimistic concurrency”, I mean a strategy that parallelizes aggressively, even if that means doing some work speculatively, based on guesses made when parallelizing the work, guesses that might prove wrong. When the guesses are wrong, the process has to spend extra time later fixing the resulting problems. But if the speedup gained by the concurrency is great enough, it can outweigh the cost of wrong guesses. (This is a bit like the way Ethernet chooses to clean up after packet collisions, instead of attempting to prevent them the way token-ring networks do.)

A fast and simple-minded process divides the XML document into chunks, and multiple parallel threads handle multiple chunks at the same time. The fragmentary results achieved by the chunkwise parsing can be knit back together quickly; sometimes the fixup process shows that one or the other chunk needs to be reparsed.

What does Moore’s Law have to do with XML parsing?

OK, so much for the elevator speech. If you’re still interested, here is a little more detail.

First, some background. Moore’s Law says, roughly, that the number of transistors it’s possible to put on a chip doubles every eighteen months. For many years, this doubling of transistor count was accompanied by increases in clock speed. (I don’t understand the connection, myself, not being an electrical engineer. I just take it on faith.) But higher clock speeds apparently require more power and generate more heat, and eventually this causes problems for the people who actually use the chips. (Very few laptop designers are persuaded that water-cooled systems can be made to have the requisite portability. And liquid nitrogen, which would be the next step? Don’t get them started.)

So nowadays the doubling of transistors appears to be reflected in the rise of multi-core chips; dual-core chips in the current crop of off-the-shelf machines, with every expectation that the number of cores on standard chips will rise. Intel has already shipped chips with four and eight cores, although I haven’t seen any four-core machines on my list when shopping for laptops. (I’m not sure whether cores are expected to double every eighteen months indefinitely, or not; if they do, will we end up with a 1024-core chip vaguely resembling the Connection Machine in our laptops in another fifteen years?)

It used to be that performance rose about as fast as the transistor count, because the clock speed kept going up; even if software didn’t do anything smarter, it kept getting faster because the chip was faster. But to get performance benefits out of multi-core chips, a system is going to want to keep all of the cores busy whenever possible. People have been thinking about parallel computing for a long time, and at the 30,000-foot level, the answer so far seems to boil down to the simple, general principle “Gosh, that’s hard!”

Under those circumstances it seems plausible that a manufacturer of multi-core chips might see it as in the manufacturer’s own interest to show people how to multi-thread common applications, so as to make it easier to get as much performance as possible out of multi-core chips.

Parallelizing parsing

How do you parallelize the task of parsing an XML input stream? (There may be other ways to keep multiple threads busy, but this seems like the obvious choice if we can manage it.)

The answer wasn’t obvious to me. There are references to parallelism in the titles or summaries of a number of papers in Grune and Jacobs’s online bibliography on parsing techniques (part of the supplemental material to their book on Parsing Techniques), but none that leap out at me as being easy to understand.

One way to parallelize XML parsing is to run a pre-scanner over the document to select suitable places to divide it into chunks, and then to parse the chunks in parallel. (There is earlier work on this by Wei Lu et al., which Yu Wu et al. cite. They also cite work on other ways to parallelize XML parsing, but I don’t understand them and won’t try to describe them.)

The problem with (the existing state of) the pre-scanning approach, according to the Intel team, is that the pre-scanning takes a lot of time by itself, and once the parsing process itself is optimized, the overhead imposed by the pre-scanner ends up being prohibitive (or at least deplorable).

Chunking

So Yu Wu et al. use a simpler (and I guess faster) pre-scanner. They don’t attempt to find wholly optimal chunk boundaries, and they don’t attempt to ensure that the parse context at the chunk boundary is completely understood by the pre-scanner. They don’t go into great detail on the chunking method, but if I have understood it correctly, the primary criteria for division of the XML document into chunks are that (1) the chunks should be of about the same size, (2) there should be at least one chunk per thread, (3) other things being equal, fewer chunks is better (because this reduces the cost of post-processing), and (4) each chunk should start with a left angle bracket.

Until I looked at the paper again just now I thought each chunk had to begin with something that looks like a start-tag, but I can’t find that constraint in the paper. Maybe I hallucinated it.

So while I don’t know the details of the Intel pre-scanner, I imagine a pre-scanner that just looks for an angle bracket in the neighborhood of the desired chunk boundary: if you have an 800-MB document you want to divide into eight chunks (in order to allocate them to eight threads, say), my imaginary pre-scanner just looks in the vicinity of byte offset 100,000,000 for a left angle bracket, scanning forwards and/or backwards until it finds one. If we are reading the document from random-access storage, the pre-scanner doesn’t even have to scan the body of the chunk.

Some readers will be saying to themselves at this point “But wait, not everything that looks like a start-tag or an end-tag is necessarily a start-tag or an end-tag. It might be in a CDATA section. It might be in a comment. What then?”

Hold that thought. You’re quite right, and we’ll come back to it.

Parallel parsing

Now, each thread is given a chunk of XML to parse, but only one thread gets to begin at the beginning of the document, so only one thread actually knows where it is. The other threads are all following the advice of the Latin poet Horace, who recommends beginning in the middle (medias in res). (I’m tempted to see if we could call this parsing technique Horatian parsing, for that reason. That’s probably not going to fly in the community at large, but it’s more compact than “hybrid parallel-processing parser” or any other descriptive phrase I can come up with denoting the algorithm presented by the Intel team, so I’ll use it in the rest of this description.)

The nature of XML, however, is that the element structure is fairly explicit in the document and doesn’t require a lot of context information. When in a chunk you see a start-tag, some content, and a matching end-tag, you know that’s an element, and you can build the appropriate data structure. But you will also run into some things that you can’t handle all by yourself: unmatched start-tags (for elements that end after the current chunk), unmatched end-tags (for elements that started in earlier chunks), and undeclared namespace prefixes (for namespace bindings declared in some ancestor element). The Horatian threads keep track of each of these phenomena, and produce as their output a convenient representation of parts of a parse tree (think DOM, not SAX), and lists of unmatched start-tags, end-tags, and namespace prefixes, together with some miscellaneous context information.

The post-processor can use the lists of unmatched bits to knit the data structures together: the last unmatched start-tag found in chunk 1 matches the first unmatched end-tag found in chunk 2, and so on.

The post-processor can also use the context information to see whether any parsing needs to be redone. (Those of you who were worried about misleading angle brackets in comments and CDATA sections and so on? Remember I told you to hold that thought? OK, put it down now. This is where we deal with it.)

If chunk n ended in the middle of a comment or in a CDATA section, then the parsing of chunk n + 1 will need to be redone. But if we have divided the document into eight chunks, and one chunk, or four, need a second parsing, we are still running at about four times the speed of a sequential parser.

Validation

Once the chunks have been knit together (and, crucially, once all namespace prefixes are bound), the same chunking is used to perform schema-validity assessment: one thread validates chunk 1, another validates chunk 2, etc.

I won’t go into the fixup needed to knit together the various partial results; it suffices to observe that in both XSD and Relax NG, the validation needed for an element cannot be determined reliably without knowing its ancestry. The consequence is that validation cannot be performed reliably without doing the post-parsing fixup first. (You could guess which element declaration to use, I suppose, but I think the rate of wrong guesses would be too high.)

If you want to pass appinfo from the governing element declaration to the consumer, you will also in pathological cases need to know the left siblings of the element, in order to know where it occurs in the content model of its parent, so you can know which appinfo to use. I expect any schema validator designed for this kind of optimistic concurrency will therefore decline to expose appinfo from element declarations.

Performance

In theory, if parallelization imposes no overhead, then by using two threads you get a two-fold speedup, four threads gets a four-fold speedup, etc.

In practice, the test results presented by the Intel group show, as one might expect, that there is some non-negligeable overhead in this scheme. For four threads, their stand-alone parser shows a slightly better than two-fold speedup; for eight threads, slightly better than four-fold. Some of the overhead, clearly, is in the pre-scanning and the post-processing, but if I’m reading their graphs correctly, for eight threads neither of these processes comes close to the amount of time spent in parsing and validation, so I’m guessing that the main performance hit comes from the need to re-parse some chunks.

Small documents (which the paper describes as those smaller than 64 KB) do not benefit from the parallelization, and larger documents (larger than 1 MB) benefit more than medium-sized ones. They didn’t talk at any length about the nature of the test data, beyond saying that it was real data from real customers.

Optimization can be a real pain sometimes, and listening to people talk about optimization can be a good cure for insomnia. But sometimes, someone has a really simple Big Idea that makes sense even to someone as ignorant as me, and then it can be an awful lot of fun to hear about it and think about the implications.

Optimistic concurrency for XML processing: a Big Idea whose implications can be huge.