Category: extremely controversial

Although with SQLite WAL we have direct-access now, we don’t support direct-access for insert and delete SPARQL queries. Those queries when made using libtracker-sparql still go over D-Bus using Adrien’s FD passing D-Bus IPC technique. The library will do that for you.

After investigating a performance analysis by somebody from Intel we learned that there is still a significant overhead per each IPC call. In the analysis the person made miner-fs combine multiple insert transactions together and then send it over as a single big transaction. This was noticeably faster than making many individual IPC requests.

The problem with this is that if one of the many insert queries fail, they all fail: not good.

We’re now experimenting with a private API that allows you to pass n individual insert transactions, and get n errors back, using one IPC call.

The numbers are promising even on Desktop D-Bus (the test):

$ cd tests/functional-tests/
$ ./update-array-performance-test
First run (first update then array)
Array: 0.103675, Update: 0.139094
Reversing run (first array then update)
Array: 0.290607, Update: 0.161749
$ ./update-array-performance-test
First run (first update then array)
Array: 0.105920, Update: 0.137554
Reversing run (first array then update)
Array: 0.118785, Update: 0.130630
$ ./update-array-performance-test
First run (first update then array)
Array: 0.108501, Update: 0.136524
Reversing run (first array then update)
Array: 0.117308, Update: 0.151192
$

We’re now deciding whether or not the API will become public; returning arrays of errors isn’t exactly ‘nice’ or ‘standard’.

While trying to handle a bug that had a description like “if I do this, tracker-store’s memory grows to 80MB and my device starts swapping”, we where surprised to learn that a sqlite3_stmt consumes about 5 kb heap. Auwch.

Before we didn’t think that those prepared statements where very large, so we threw all of them in a hashtable for in case the query was ran again later. However, if you collect thousands of such statements, memory consumption obviously grows.

We decided to implement a LRU cache for these prepared statements. For clients that access the database using direct-access the cache will be smaller, so that max consumption is only a few megabytes. Because our INSERT and DELETE queries are more reusable than SELECT queries, we split it into two different caches per thread.

The implementation is done with a simple intrinsic linked ring list. We’re still testing it a little bit to get good cache-size numbers. I guess it’ll go in master soon. For your testing pleasure you can find the branch here.

The crawler’s modification time queries

Yesterday we optimized the crawler’s query that gets the modification time of files. We use this timestamp to know whether or not a file must be reindexed.

Originally, we used a custom SQLite function called tracker:uri-is-parent() in SPARQL. This, however, caused a full table scan. As long as your SQL table for nfo:FileDataObjects wasn’t too large, that wasn’t a huge problem. But it didn’t scale linear. I started with optimizing the function itself. It was using a strlen() so I replaced that with a sqlite3_value_bytes(). We only store UTF-8, so that worked fine. It gained me ~ 10%; not enough.

So this commit was a better improvement. First it makes nfo:belongsToContainer an indexed property. The x nfo:belongsToContainer p means x is in a directory p for file resources. The commit changes the query to use the property that is now indexed.

The original query before we started with this optimization took 1.090s when you had ~ 300,000 nfo:FileDataObject resources. The new query takes about 0.090s. It’s of course an unfair comparison because now we use an indexed property. Adding the index only took a total of 10s for a ~ 300,000 large table and the table is being queried while we index (while we insert into it). Do the math, it’s a huge win in all situations. For the SQLite freaks; the SQLite database grew by 4 MB, with all items in the table indexed.

PDF extractor

Another optimization I did earlier was the PDF extractor. Originally, we used the poppler-glib library. This library doesn’t allow us to set the OutputDev at runtime. If compiled with Cairo, the OutputDev is in some versions a CairoOutputDev. We don’t want all images in the PDF to be rendered to a Cairo surface. So I ported this back to C++ and made it always use a TextOutputDev instead. In poppler-glib master this appears to have improved (in git master poppler_page_get_text_page is always using a TextOutputDev).

Another major problem with poppler-glib is the huge amount of copying strings in heap. The performance to extract metadata and content text for a 70 page PDF document without any images went from 1.050s to 0.550s. A lot of it was caused by copying strings and GValue boxing due to GObject properties.

Table locked problem

Last week I improved D-Bus marshaling by using a database cursor. I forgot to handle SQLITE_LOCKED while Jürg and Carlos had been introducing multithreaded SELECT support. Not good. I fixed this; it was causing random Table locked errors.

Before

For returning the results of a SPARQL SELECT query we used to have a callback like this. I removed error handling, you can find the original here.

We need to marshal a database result_set to a GPtrArray because dbus-glib fancies that. This is a lot of boxing the strings into GValue and GStrv. It does allocations, so not good.

static void
query_callback(TrackerDBResultSet *result_set,GError *error,gpointer user_data)
{
  TrackerDBusMethodInfo *info = user_data;
  GPtrArray *values = tracker_dbus_query_result_to_ptr_array (result_set);
  dbus_g_method_return (info->context, values);
  tracker_dbus_results_ptr_array_free (&values);
}

void
tracker_resources_sparql_query (TrackerResources *self, const gchar *query,
                                DBusGMethodInvocation *context, GError **error)
{
  TrackerDBusMethodInfo *info = ...; guint request_id;
  TrackerResourcesPrivate *priv= ...; gchar *sender;
  info->context = context;
  tracker_store_sparql_query (query, TRACKER_STORE_PRIORITY_HIGH,
                              query_callback, ...,
                              info, destroy_method_info);
}

After

Last week I changed the asynchronous callback to return a database cursor. In SQLite that means an sqlite3_step(). SQLite returns const pointers to the data in the cell with its sqlite3_column_* APIs.

This means that now we’re not even copying the strings out of SQLite. Instead, we’re using them as const to fill in a raw DBusMessage:

static void
query_callback(TrackerDBCursor *cursor,GError *error,gpointer user_data)
{
  TrackerDBusMethodInfo *info = user_data;
  DBusMessage *reply; DBusMessageIter iter, rows_iter;
  guint cols; guint length = 0;
  reply = dbus_g_method_get_reply (info->context);
  dbus_message_iter_init_append (reply, &iter);
  cols = tracker_db_cursor_get_n_columns (cursor);
  dbus_message_iter_open_container (&iter, DBUS_TYPE_ARRAY,
                                    "as", &rows_iter);
  while (tracker_db_cursor_iter_next (cursor, NULL)) {
    DBusMessageIter cols_iter; guint i;
    dbus_message_iter_open_container (&rows_iter, DBUS_TYPE_ARRAY,
                                      "s", &cols_iter);
    for (i = 0; i < cols; i++, length++) {
      const gchar *result_str = tracker_db_cursor_get_string (cursor, i);
      dbus_message_iter_append_basic (&cols_iter,
                                      DBUS_TYPE_STRING,
                                      &result_str);
    }
    dbus_message_iter_close_container (&rows_iter, &cols_iter);
  }
  dbus_message_iter_close_container (&iter, &rows_iter);
  dbus_g_method_send_reply (info->context, reply);
}

Results

The test is a query on 13500 resources where we ask for two strings, repeated eleven times. I removed a first repeat from each round, because the first time the sqlite3_stmt still has to be created. This means that our measurement would get a few more milliseconds. I also directed the standard out to /dev/null to avoid the overhead created by the terminal. The results you see below are the value for “real”.

There is of course an overhead created by the “tracker-sparql” program. It does demarshaling using normal dbus-glib. If your application uses DBusMessage directly, then it can avoid the same overhead. But since for both rounds I used the same “tracker-sparql” it doesn’t matter for the measurement.

$ time tracker-sparql -q "SELECT ?u  ?m { ?u a rdfs:Resource ;
          tracker:modified ?m }" > /dev/null

Without the optimization:

0.361s, 0.399s, 0.327s, 0.355s, 0.340s, 0.377s, 0.346s, 0.380s, 0.381s, 0.393s, 0.345s

With the optimization:

0.279s, 0.271s, 0.305s, 0.296s, 0.295s, 0.294s, 0.295s, 0.244s, 0.289s, 0.237s, 0.307s

The improvement ranges between 7% and 40% with average improvement of 22%.

Every (good) developer knows that copying of memory and boxing, especially when dealing with a large amount of pieces like members of collections and the cells in a table, are a bad thing for your performance.

More experienced developers also know that novice developers tend to focus on just their algorithms to improve performance, while often the single biggest bottleneck is needless boxing and allocating. Experienced developers come up with algorithms that avoid boxing and copying; they master clever pragmatical engineering and know how to improve algorithms. A lot of newcomers use virtual machines and script languages that are terrible at giving you the tools to control this and then they start endless religious debates about how great their programming language is (as if it matters). (Anti-.NET people don’t get on your horses too soon: if you know what you are doing, C# is actually quite good here).

We were of course doing some silly copying ourselves. Apparently it had a significant impact on performance.

Once Jürg and Carlos have finished the work on parallelizing SELECT queries we plan to let the code that walks the SQLite statement fill in the DBusMessage directly without any memory copying or boxing (for marshalling to DBus). We found the get_reply and send_reply functions; they sound useful for this purpose.

I still don’t really like DBus as IPC for data transfer of Tracker’s RDF store’s query results. Personally I think I would go for a custom Unix socket here. But Jürg so far isn’t convinced. Admittedly he’s probably right; he’s always right. Still, DBus to me doesn’t feel like a good IPC for this data transfer..

We know about the requests to have direct access to the SQLite database from your own process. I explained in the bug that SQLite3 isn’t MVCC and that this means that your process will often get blocked for a long time on our transaction. A longer time than any IPC overhead takes.