<html>
<body>
<div style="font-family: Verdana, Arial, Helvetica, Sans-Serif;">
<table bgcolor="#f9f3c9" width="100%" cellpadding="8" style="border: 1px #c9c399 solid;">
<tr>
<td>
This is an automatically generated e-mail. To reply, visit:
<a href="http://git.reviewboard.kde.org/r/113355/">http://git.reviewboard.kde.org/r/113355/</a>
</td>
</tr>
</table>
<br />
<table bgcolor="#fefadf" width="100%" cellspacing="0" cellpadding="8" style="background-image: url('http://git.reviewboard.kde.org/static/rb/images/review_request_box_top_bg.ab6f3b1072c9.png'); background-position: left top; background-repeat: repeat-x; border: 1px black solid;">
<tr>
<td>
<div>Review request for kdelibs.</div>
<div>By Frank Reininghaus.</div>
<p style="color: grey;"><i>Updated Oct. 26, 2013, 5:56 p.m.</i></p>
<h1 style="color: #575012; font-size: 10pt; margin-top: 1.5em;">Changes</h1>
<table width="100%" bgcolor="#ffffff" cellspacing="0" cellpadding="10" style="border: 1px solid #b8b5a0">
<tr>
<td>
<pre style="margin: 0; padding: 0; white-space: pre-wrap; white-space: -moz-pre-wrap; white-space: -pre-wrap; white-space: -o-pre-wrap; word-wrap: break-word;">The main changes in this version are:
1. Addressed Mark's comments - shame on me for re-implementing the toList() function ;-)
2. Brought back the UDSEntry::reserve(int) function, which allows to save some time if kioslaves call it before filling the entry.
3. Modified the benchmark so that the number of fields can be changed easily.
To see how much we could possibly lose by dropping the O(1) access of QHash, I benchmarked also 70 fields (there are currently 30 different keys defined in UDSEntry, plus an extra space of 40 between UDS_EXTRA and UDS_EXTRA_END).
The results in millisecs are for
master/r4 of this patch (with the "reserve", which can save time, disabled in "createEntries"):
createEntries: 3307/2841
saveEntries: 1349/972
loadEntries: 2394/1339
readFieldsFromEntries: 2211/2425
So there is a slowdown when reading the fields, but it's not really big.</pre>
</td>
</tr>
</table>
<div style="margin-top: 1.5em;">
<b style="color: #575012; font-size: 10pt;">Repository: </b>
kdelibs
</div>
<h1 style="color: #575012; font-size: 10pt; margin-top: 1.5em;">Description </h1>
<table width="100%" bgcolor="#ffffff" cellspacing="0" cellpadding="10" style="border: 1px solid #b8b5a0">
<tr>
<td>
<pre style="margin: 0; padding: 0; white-space: pre-wrap; white-space: -moz-pre-wrap; white-space: -pre-wrap; white-space: -o-pre-wrap; word-wrap: break-word;">This patch is based on some discussions on the kfm-devel mailing list, see http://lists.kde.org/?t=137935784800003&r=1&w=2
Mark found out that KIO's UDSEntry class is one of the major consumers of memory in applications which use KIO to list directories with a large number of files, and I found a way use implicit sharing to drastically reduce the amount of memory it needs. Many thanks to Milian for his great blog post http://milianw.de/blog/katekdevelop-sprint-vienna-2012-take-1, without which I would probably not have had such ideas.
1. The problem
The UDSEntry keeps all sorts of information about files which can be stored in a string (name, user, group, etc.) or in a long long (file size, modification time, etc.). All these data can be accessed with a uint key, and UDSEntry returns the corresponding QString or long long in O(1) time. Internally, it stores the data in a QHash<uint, Field>, where Field is a struct that has a QString and a long long member.
The problem is that QHash needs quite a lot of memory to provide the O(1) access, see http://doc.qt.digia.com/qq/qq19-containers.html for details, and that the minimum capacity of a QHash seems to be 17, even though the number of entries in a UDSEntry is often 8 in the rather typical standard kio_file use case.
2. Proposed solution
(a) We can store the "Fields" in a QVector<Field>, which needs only very little overhead compared to the memory that the actual "Fields" need. When loading a UDSEntry from a QDataStream, we just append all "Fields" to this QVector one by one. Moreover, we need a QHash<uint, int>, which maps each key to the index of its Field in the QVector. This restructuring alone does not reduce the memory usage, of course.
(b) The key observation is that the QDataStream, which KIO::ListJob reads the UDSEntries from, typically provides the different "Fields" in exactly the same order. This means that the QHash<uint, int> is usually exactly the same for every UDSEntry, and we can make use of implicit sharing to store only one copy of this QHash. I've modified
void UDSEntryPrivate::load(QDataStream &s, UDSEntry &a)
such that it remembers the most recent QHash<uint, int> and just adds an implicitly shared copy of it to "a" if the order of the Fields has not changed.
(c) Moreover, some of the QString Fields in the UDSEntries in one directory are often the same, like, e.g., the user and the group. The "load" function also remembers the most recently read values for each Field in a static QVector<QString> and just puts an implicitly shared copy into the UDSEntry if possible.
3. Possible disadvantages
(a) When using the "remove" member, the new version of UDSEntry does not remove the Field from the QVector<Field>. This means that removing and adding a "Field" repeatedly would let the memory usage grow indefinitely. According to David (http://lists.kde.org/?l=kfm-devel&m=138052519927973&w=2), this doesn't matter though because no known user of UDSEntry uses its remove() member. Maybe we should remove remove (pun stolen grom David) in the frameworks branch then?
(b) In principle, the old version of UDSEntryPrivate::load(QDataStream&, UDSEntry&) was reentrant. This is not the case for my changed version. Reentrancy could be restored rather easily by protecting the access to the static data with a mutex, but given that most of KIO is not supposed to be used from outside the main thread AFAIK, I don't know if this is necessary.
4. Changes since the first version of the patch which I posted in http://lists.kde.org/?l=kfm-devel&m=137962995805432&w=2
(a) Implemented the minor changes suggested by David in http://lists.kde.org/?l=kfm-devel&m=137975442807965&w=2
(b) Added a unit test to verify that storing and loading UDSEntries from a stream works even if the order of the fields is permuted, and some fields are removed or added in between.
(c) Fixed a bug which was uncovered by the test: cachedUdsFields.erase(cachedUdsFields.begin() + i, cachedUdsFields.end()) instead of cachedUdsFields.erase(cachedUdsFields.begin() + i)
(d) Use QVector::reserve to reserve the appropriate size for the QVector<Field>. Saves some time when loading the UDSEntry, and reduces the memory usage further.
(e) Changed the type of the loop variable from quint32 to int to fix some compiler warnings.
</pre>
</td>
</tr>
</table>
<h1 style="color: #575012; font-size: 10pt; margin-top: 1.5em;">Testing </h1>
<table width="100%" bgcolor="#ffffff" cellspacing="0" cellpadding="10" style="border: 1px solid #b8b5a0">
<tr>
<td>
<pre style="margin: 0; padding: 0; white-space: pre-wrap; white-space: -moz-pre-wrap; white-space: -pre-wrap; white-space: -o-pre-wrap; word-wrap: break-word;">Old and new unit tests pass. The memory usage of Dolphin when loading a directory with 100,000 files in Icons View is reduced from 165.4 MB to 113.3 MB. Any application that uses a file dialog, a KDirLister, or anything else that uses a KIO::ListJob to list directory contents should benefit from similar savings.</pre>
</td>
</tr>
</table>
<h1 style="color: #575012; font-size: 10pt; margin-top: 1.5em;">Diffs</b> (updated)</h1>
<ul style="margin-left: 3em; padding-left: 0;">
<li>kio/kio/udsentry.h <span style="color: grey">(e1c8b05)</span></li>
<li>kio/kio/udsentry.cpp <span style="color: grey">(1e1f503)</span></li>
<li>kio/tests/CMakeLists.txt <span style="color: grey">(1019312)</span></li>
<li>kio/tests/simplelistjobtest.cpp <span style="color: grey">(PRE-CREATION)</span></li>
<li>kio/tests/udsentrybenchmark.cpp <span style="color: grey">(PRE-CREATION)</span></li>
<li>kio/tests/udsentrytest.h <span style="color: grey">(PRE-CREATION)</span></li>
<li>kio/tests/udsentrytest.cpp <span style="color: grey">(PRE-CREATION)</span></li>
</ul>
<p><a href="http://git.reviewboard.kde.org/r/113355/diff/" style="margin-left: 3em;">View Diff</a></p>
</td>
</tr>
</table>
</div>
</body>
</html>