Dense Meta Data using MMap-Based arrays #80

joka921 · 2018-07-31T09:14:18Z

There is still much to do and to refactor, just wanted a first chance for reviewing.

niklas88 · 2018-08-01T09:58:12Z

src/index/CMakeLists.txt

@@ -5,10 +5,13 @@ add_library(index
        VocabularyGenerator.h VocabularyGenerator.cpp
        ConstantsIndexCreation.h
        ExternalVocabulary.h ExternalVocabulary.cpp
-        IndexMetaData.h IndexMetaData.cpp
+	IndexMetaData.h IndexMetaDataImpl.h IndexMetaDataImpl.cpp


This indentation looks wrong

niklas88 · 2018-08-01T09:58:26Z

src/index/CMakeLists.txt

        StxxlSortFunctors.h
        TextMetaData.cpp TextMetaData.h
        DocsDB.cpp DocsDB.h
+	MmapBasedArray.h MmapBasedArrayImpl.h


This indentation looks wrong, you might want to set your editor to show spaces and tabs differently

niklas88 · 2018-08-01T10:01:07Z

src/index/Index.cpp

@@ -216,6 +220,11 @@ size_t Index::passTsvFileForVocabulary(const string& tsvFile) {
  }
  LOG(INFO) << "Pass done.\n";
  _vocab.createFromSet(items);
+  // TODO: completely merge TSV and NTrIPLEs parsing. This is just for making


TODO(<username>)

niklas88 · 2018-08-01T10:03:22Z

src/index/Index.cpp

@@ -885,24 +896,19 @@ void Index::createFromOnDiskIndex(const string& onDiskBase,
  setOnDiskBase(onDiskBase);
  _vocab.readFromFile(_onDiskBase + ".vocabulary",
                      _onDiskLiterals ? _onDiskBase + ".literals-index" : "");
+  // this is the read only version, TODO: give better naming or type


TODO(<username>)

niklas88 · 2018-08-01T10:04:20Z

src/index/Index.cpp

@@ -885,24 +896,19 @@ void Index::createFromOnDiskIndex(const string& onDiskBase,
  setOnDiskBase(onDiskBase);
  _vocab.readFromFile(_onDiskBase + ".vocabulary",
                      _onDiskLiterals ? _onDiskBase + ".literals-index" : "");
+  // this is the read only version, TODO: give better naming or type
+  _opsMeta.setup(onDiskBase + MMAP_FILE_SUFFIX + ".ops");


Isn't setup() basically open() so could be named open()?

Well, open() would suggest that we open a file, but actually what we are doing is completely depending on our implementation (nothing, reloading a mmapBased array, copying something to a stxxl vector (deprecated) etc.
My purpose was to create a generic function where the different template instatiations can have different initializations/setups passed through by varargs. I chose "setup" because you anyway have to look at the concrete type and which type of setup is required.

Sounds reasonable

niklas88 · 2018-08-01T10:27:16Z

src/index/MmapBasedArray.h

@@ -0,0 +1,190 @@
+// Copyright 2018, University of Freiburg,


I think this belongs in util because it is quite generic. Also it should be namespaced and "Based" is redundant so just name it MmapArray or better yet since it can grow MmapVector

niklas88 · 2018-08-01T10:28:15Z

src/index/MmapBasedArray.h

+    AD_CHECK(idx < _size);
+  }
+
+  // we will never hold less than this capacity


"never have less capacity" as I guess we can hold 0 items

niklas88 · 2018-08-01T10:29:55Z

src/index/MmapBasedArray.h

+  // create Array of given size  fill with default value
+  // file contents will be overriden if existing!
+  // allows read and write access
+  void setupNewFile(size_t size, const T& defaultValue, string filename);


setupNewFile() sounds like open() to me

In this case I agree, I would refactor it using an argument of type std::ios_base::openmode and mimic the read/write semantics used with ifstreams.

I decided to use two "tag"-classes. You will see if you like it better that way.

niklas88 · 2018-08-01T10:38:17Z

src/index/MmapBasedArrayImpl.h

+  _bytesize = realSize._bytesize;
+  // writeMetaData will also adapt the filesize according to
+  // _capacity and _bytesize
+  unmap();


On Linux we should probably use mremap()

niklas88 · 2018-08-01T10:38:44Z

src/index/VocabularyGenerator.h

-// Writes file "externalTextFile" which can be used to directly write external
-// Literals
-void mergeVocabulary(const std::string& basename, size_t numFiles);
+// Writes file "externalTextFile" which can be used to directly write ext


You seem to have deleted the rest of the line by accident

niklas88 · 2018-08-03T14:10:58Z

src/util/MmapVector.h

+    size_t _bytesize;
+  };
+
+  // 2 tags two differentiate different versions of the 


niklas88

Couple more small comments. Also I think we should have unit tests for the new stuff. That said I'll now give this version a shot with the Clueweb Freebase instance and see how it fares.

EDIT: I forgot this needs an index rebuild. How hard would it be to do automatic conversion? This would also be a nice use of the magic number + version data

niklas88 · 2018-08-03T14:12:16Z

src/util/MmapVector.h

+    T& at(size_t idx) {
+      throwIfUninitialized();
+      if (idx >= _size) {
+	throw std::out_of_range();


missing indent

niklas88 · 2018-08-03T14:12:33Z

src/util/MmapVector.h

+    const T& at(size_t idx) const {
+      throwIfUninitialized();
+      if (idx >= _size) {
+	throw std::out_of_range();


missing indent

niklas88 · 2018-08-03T14:19:07Z

src/util/MmapVector.h

+    MmapVector& operator=(const MmapVector<T>&) = delete;
+
+    // move construction and assignment 
+    // TODO<joka921>: is this noexcept? 


I think so, yes

niklas88 · 2018-08-03T14:20:24Z

src/util/MmapVector.h

+    MmapVector(MmapVector<T>&& other);
+    MmapVector& operator=(MmapVector<T>&& other);
+
+    // Construct from anything that has a valid open() method


It doen't really construct from something else. It just constructs and opens in one go

niklas88 · 2018-08-03T14:22:20Z

src/util/MmapVector.h

+    // truncate the underlying file to _bytesize and write meta data (_size,
+    // _capacity, etc) for this
+    // array to the end. new size of file will be _bytesize + sizeof(meta data)
+    // TODO: maybe extend meta data by "magic number" constant to verify if this


make it a magic number + vesion

niklas88 · 2018-08-03T15:03:51Z

src/index/MetaDataHandler.h

+  }
+
+  // __________________________________________________________
+  const FullRelationMetaData& getAsserted(Id id) const {


what does "Asserted" mean here?

It means that there is a AD_CHECK whether there is a valid entry at this Id. That is how the original HashMap based code used it and so I chose to keep this for the other implementations as well.
The closest name from the stl would be "at", would you prefer that?

niklas88 · 2018-08-03T15:08:25Z

src/index/MetaDataHandler.h

+
+// _____________________________________________________________
+// hidden in implementation namespace
+namespace ExtVecWrapperImpl {


It's not really important that ExtVec is external, right? So why not just call it Vec?

niklas88 · 2018-08-03T15:17:54Z

src/index/MetaDataTypes.h

+  uint64_t _typeMultAndNofElements;
+};
+
+inline ad_utility::File& operator<<(ad_utility::File& f,


I see, makes sense

niklas88 · 2018-08-03T15:19:21Z

src/index/MetaDataTypes.h

+  return f;
+}
+
+class BlockBasedRelationMetaData {


Hmm, if you want to give it a shot I think you can only improve upon the current state but I guess it's not strictly necessary for the PR

niklas88 · 2018-08-03T15:20:38Z

src/util/MmapVector.h

+
+  // 2 tags two differentiate different versions of the 
+  // setup / construction variants of Mmap Vector which only take a filename
+  class createTag{};


These are still classes so should probably start with a capital letter.

niklas88 · 2018-08-03T18:19:36Z

src/util/MmapVectorImpl.h

@@ -35,19 +41,35 @@ void MmapVector<T>::writeMetaDataToEnd() {
  ofs.write((char*)&_size, sizeof(_size));
  ofs.write((char*)&_capacity, sizeof(_capacity));
  ofs.write((char*)&_bytesize, sizeof(_bytesize));
+  ofs.write((char*)&MagicNumber, sizeof(MagicNumber));


since magic number is larger than an unit8_t this means the written value will depend on the endianess. But I think in this case this might not be bad since we can't read other-endianess data anyway since all of it is written in the non--endianess-portable write(&data, sizeof(data)); way.

niklas88 · 2018-08-03T18:21:30Z

src/util/MmapVector.h

-  static constexpr float ResizeFactor = 1.5;
-  static constexpr int MagicNumber = 7601577;
-  static constexpr int Version = 0;
+ private:

 private:


One of the private: is redundant

niklas88 · 2018-08-03T18:23:50Z

Do you already have some experience with how much this reduces RAM usage and startup time on one of our larger datasets?

niklas88 · 2018-08-06T12:19:03Z

src/index/MetaDataConverter.cpp

+
+  if (verify) {
+    for (auto it = res._data.cbegin(); it != res._data.cend(); ++it) {
+      if (hmap._data.getAsserted((*it).first) != (*it).second) {


Any reason you use (*it).first instead of it->first? I feel like the second is more readable

My iterator manually "plunges" the index to its pointee, so we get a map-like interface where
*it is a pair of <key, value>. There are some restrictions to the implementation of operator-> which must yield something that is or behaves as a pointer. I can check if I can make this work, but there were some issues.

I have now implemented a variant of operator-> for you. I am not sure if this is the most elegant or efficient way with the std::reference_wrappers, but during our uses we normally don't use it for queries.

niklas88 · 2018-08-06T12:24:07Z

src/index/IndexMetaData.cpp

+  size_t nofBytesDone = sizeof(size_t);
+  _name.assign(reinterpret_cast<char*>(buf + nofBytesDone), nameLength);
+  nofBytesDone += nameLength;
+  // skip nOfRelations, since this information is already stored in


Why are we sure there were previous runs?

because we first have to run the IndexBuilder to create MetaData which is then written to disk which we then read again when opening the server using createFromByteBuffer.
The local nofRelations variable is needed in the non-persistent specializations of IndexMetaData where we read the FullRelationalMetaData from our byteBuffer. (You can check the difference to the "standard" implementation.
I also added a (hopefully) better comment to this.

niklas88 · 2018-08-06T12:47:08Z

src/index/Index.cpp

-    _spoFile.read(buf, static_cast<size_t>(metaTo - metaFrom), metaFrom);
-    _spoMeta.createFromByteBuffer(buf);
-    delete[] buf;
+    _spoMeta.readFromFile(&_spoFile);


This should also check for a denseMetaFile for extremely RAM saving configs right?

Well, that is a relatively deep design decision:
Currently:
P** and S** are always "sparse" (hashMapBased on disk), while O** is dense.
(hardcoded using templates).

If we want to have this at runtime (decide during index creation which meta data we want for which format) I think we would have to let the different IndexMetaData inherit from a (virtual) class IndexMetaDataBase. Then we could use a unique_ptr to dispatch at runtime. The last time I mentioned it I thing you were not really in favor of that, but it would make things more flexible at the cost of the pointer indirections.

But I think I would first finish this current PR and then think about this additional flexibility which would change quite some lines in the Index class.

Ahh yeah I remember, ok I think for now we should keep the current approach.

niklas88 · 2018-08-06T13:05:29Z

src/MetaDataConverterMain.cpp

+      exit(1);
+    }
+    std::string base = in.substr(0, in.size() - 4);
+    out = base + DENSE_META_SUFFIX + suffix;


This suffix seems to be only used when using the converter. I think a converted index should have the exact same data as the same index built freshly in the files named the same (unless both versions have a version field and we can check that in the code).

As I read from your Email you would prefer to always write the dense meta data to an extra file. I will change this.

(together with my next PR for prefix compression I am preparing a index.configuration file (json) where we can also save information about the index in a more explicit way.

@joka921 it doesn't really matter if it's one file. It's just that an index upgrade tool like the new MetaDataConvertor should output the same format as would be used when building from scratch with that version.

niklas88 · 2018-08-06T13:52:12Z

Somehow I'm getting a compile error when building this branch locally using

git checkout -b joka921-f.mmapBasedArray master

git pull git://github.com/joka921/QLever.git f.mmapBasedArray

Looks like the linker can't find the specialization's code

CMakeFiles/IndexMetaDataTest.dir/IndexMetaDataTest.cpp.o: In function `IndexMetaDataTest_writeReadTest2Mmap_Test::TestBody()':
IndexMetaDataTest.cpp:(.text+0x4cd0): undefined reference to `ad_utility::File& operator<<     <MetaDataWrapperDense<ad_utility::MmapVector<FullRelationMetaData> > >(ad_utility::File&,         IndexMetaData<MetaDataWrapperDense<ad_utility::MmapVector<FullRelationMetaData> > > const&)'
collect2: error: ld returned 1 exit status
make[2]: *** [test/CMakeFiles/IndexMetaDataTest.dir/build.make:89: test/IndexMetaDataTest] Error 1
make[1]: *** [CMakeFiles/Makefile2:2852: test/CMakeFiles/IndexMetaDataTest.dir/all] Error 2
make: *** [Makefile:163: all] Error 2

Can't find this error on Travis though but it happens on both my local Arch Linux workstation as well as on a Ubuntu 16.04 server.

joka921 · 2018-08-06T14:02:41Z

@niklas88 Have you tried running make clean before? I have the impression that somewhere cmake does not handle all dependencies right.

niklas88 · 2018-08-06T14:08:20Z

@joka921 yes and I even built with a completely fresh ./build before that on both machines

niklas88 · 2018-08-06T14:16:47Z

@joka921 ok found the problem. Removing inline in front of that function in src/IndexMetaData.cpp fixes the build. I guess the difference was that I was building with optimizations and thus this was actually inlined, removing the copy to link to. On Travis on the other hand things are built without optimization and thus a copy of the method ends up in the object.

niklas88

Hmm, I'm still unsure about these file format changes. I think we should make sure we get this right. I actually liked having the meta data in the same file. However in the previous version if I got this right, the converter would do something different than building the index from scratch, which I think is quite fragile. Namely it would create external meta data files for what was just written at the end in the freshly built case.

This difference is what I didn't like. Always creating external meta data of course fixes this but also seems a bit clumsy. Also I think that just as the MmapVector adds a magic number and version field I think it would be nice to have that in the meta data for "*." files as well. Maybe with a magic number we can then also keep them single files with the meta data at the end?

niklas88 · 2018-08-07T08:33:37Z

src/index/Index.cpp


-  metaData.appendToFile(&out);
+  // metaData.appendToFile(&out);


niklas88 · 2018-08-07T08:34:55Z

src/index/Index.cpp

-  _psoFile.open(string(_onDiskBase + ".index.pso").c_str(), "r");
-  _posFile.open(string(_onDiskBase + ".index.pos").c_str(), "r");
+  auto psoName = string(_onDiskBase + ".index.pso");
+  _psoFile.open(psoName.c_str(), "r");


I think it's time we add an open(const std::string&) method to ad_utility::File

niklas88 · 2018-08-07T08:56:30Z

src/MetaDataConverterMain.cpp

@@ -25,7 +25,7 @@ int main (int argc, char** argv) {
      exit(1);
    }
    std::string base = in.substr(0, in.size() - 4);
-    out = base + DENSE_META_SUFFIX + suffix;
+    out = in + DENSE_META_SUFFIX;


Looks like this gives us the dense file with ".ops.meta-dense" while below we get ".meta-mmap.ops".
I think since all of this changes the Index format we really should make sure to get this right. I actually liked having the meta data in the same file, my main concern in the last review was that the Converter should create the same format as a newly built index. Maybe since this is the first time we add a magic number and version we should in fact keep the meta data at the end of the files and only change the converter so it works for sparse meta data as well making sure all the files get a version and magic number.

niklas88 · 2018-08-07T09:00:37Z

src/index/IndexMetaData.h


+  // already Existing means we do not need to at rmd to _data


niklas88 · 2018-08-07T09:01:21Z

src/index/IndexMetaData.h


+  // already Existing means we do not need to at rmd to _data
+  // if bool alreadyExisting is set, we will not add argument rmd
+  // but assume that is already is contained in _data (for persistent metaData


"that it is already contained"

niklas88 · 2018-08-07T14:12:15Z

As discussed on the phone, we should also add a SPARQL query that uses one of the OSP/OPS indices to the e2e tests.

The following example query for all predicates connecting two scientists does a full scan of OSP

SELECT DISTINCT ?p WHERE {
    ?x <is-a> <Scientist> .
    ?y <is-a> <Scientist> .
    ?x ?p ?y .
}

niklas88

Looks good, any idea if the advise(AccessPattern::Sequential) calls make building the vector faster?

niklas88 · 2018-08-07T17:29:11Z

src/index/Index.cpp

@@ -922,9 +922,9 @@ void Index::createFromOnDiskIndex(const string& onDiskBase,
            << std::endl;
  if (allPermutations) {
    _opsMeta.setup(onDiskBase + ".index.ops" + MMAP_FILE_SUFFIX,
-                   ad_utility::ReuseTag());
+                   ad_utility::ReuseTag(), ad_utility::AccessPattern::Random);


Am I seeing this right, ReuseTag needs to be a tag to select at compile time but AccessPattern can be an enum because it selects at runtime?

If you want to I can also make ReuseTag an enum. It is basically the same as in an fstream where we also can decide by a runtime argument wheter we want to overwrite an existing file or use/add to its contents. The efficiency difference would not matter and maybe it would be more consistent. Just tell me, what you think.

niklas88 · 2018-08-07T17:31:25Z

src/util/MmapVector.h

@@ -224,11 +241,15 @@ class MmapVector {
  // invalidates iterators
  void unmap();

+  // advise the kernel to use a certain access pattern
+  void advise(AccessPattern pattern);


what's the difference between advice() and setAccessPattern() seems a bit arbitrary

niklas88 · 2018-08-07T17:32:19Z

src/util/MmapVectorImpl.h

  for (size_t i = 0; i < _size; ++i) {
    _ptr[i] = defaultValue;
  }
+  advise(_pattern);


interesting, so you readvise when changing the access pattern

Did this answer your question? advise is the private implementation detail (call to madvise) and setAccessPattern the global interface, that let's us for example first advise to use sequential for insertion and then switch to random for access etc...
Concerning your question about the effect: According to the manpages the "sequential" mode does "heavy prefetching" and also evicts pages soon after they are written. I did not yet measure it and think this heavily depends on the overall system load and the kernel implementation (there is a reason why it's called "advise")

niklas88 · 2018-08-07T17:34:18Z

src/MetaDataConverterMain.cpp

-      std::cerr << "error, infile must have suffix of"
-                   ".ops, .osp\n";
-      exit(1);
+  std::vector<std::string> sparseNames{".pso", ".pos", ".spo", ".sop"};


std::array

niklas88 · 2018-08-07T17:34:34Z

src/MetaDataConverterMain.cpp

+                                              permutName + ".converted");
+  }
+
+  std::vector<std::string> denseNames{".osp", ".ops"};


std::array

niklas88 · 2018-08-07T17:34:53Z

src/MetaDataConverterMain.cpp


 // _________________________________________________________
 int main (int argc, char** argv) {
  if (argc != 2) {
-    std::cerr << "Usage: ./MetaDataConverterMain <originalFilename>\n";
+    std::cerr << "Usage: ./MetaDataConverterMain <pathToIndex>\n";


I think we call this an "index prefix" elsewhere

niklas88 · 2018-08-07T17:35:11Z

src/global/Constants.h

@@ -51,6 +51,8 @@ static const int DEFAULT_NOF_VALUE_MANTISSA_DIGITS = 30;
 static const int DEFAULT_NOF_DATE_YEAR_DIGITS = 19;

 static const std::string MMAP_FILE_SUFFIX = ".meta-mmap";
+/*


remove commented

niklas88 · 2018-08-10T13:56:58Z

I think the commit you squashed this to now needs a better name

joka921 · 2018-08-10T17:30:22Z

@niklas88 This is now a feature-fixed (I hope) version. I hope we can merge it early next week so other projects can go on.

niklas88 · 2018-08-11T08:54:09Z

Can you add just another commit to update the README.md with the conversion tool and also the last section about memory use?

joka921 · 2018-08-11T14:05:10Z

@niklas88 I have added the requested documentation. After you've had a look at it I would rebase it into the actual implementation commit to get a cleaner history.

niklas88

Looks great, really happy how this turned out! Some final comments on a few typos/formatting and then feel free to merge!

niklas88 · 2018-08-11T19:28:06Z

README.md

@@ -380,6 +380,25 @@ Group by is supported, but aggregate aliases may currently only be used within t
 Supported aggregates are `MIN, MAX, AVG, GROUP_CONCAT, SAMPLE, COUNT, SUM`. All of the aggreagates support `DISTINCT`, e.g. `(GROUP_CONCAT(DISTINCT ?a) as ?b)`.
 Group concat also supports a custom separator: `(GROUP_CONCAT(?a ; separator=" ; ") as ?concat)`. Xsd types float, decimal and integer are recognized as numbers, other types or unbound variables (e.g. no entries for an optional part) in one of the aggregates that need to interpret the variable (e.g. AVG) lead to either no result or nan. MAX with an unbound variable will always return the unbound variable.

+# 6. Converting Old Indices For Current QLever Versions
+
+We have recently updated the way the index meta data (offsets of relations


Please format these lines to roughly be of equal length (up to 80 columns). When using vim this can be done with first selecting them <esc><V><j..> and then doing <g><q>. Other than that this looks great.

niklas88 · 2018-08-11T19:29:04Z

src/index/CMakeLists.txt

        StxxlSortFunctors.h
        TextMetaData.cpp TextMetaData.h
        DocsDB.cpp DocsDB.h
        FTSAlgorithms.cpp FTSAlgorithms.h)

 target_link_libraries(index parser ${STXXL_LIBRARIES})
+
+add_library(metaConverter
+        MetaDataConverter.cpp MetaDataConverter.h)


this is wrongly indented

niklas88 · 2018-08-11T19:32:22Z

src/index/Index.cpp

+                   ad_utility::ReuseTag(), ad_utility::AccessPattern::Random);
+    _ospMeta.setup(onDiskBase + ".index.osp" + MMAP_FILE_SUFFIX,
+                   ad_utility::ReuseTag(), ad_utility::AccessPattern::Random);
+    // TODO<joka921>: Refactor (upcoming PR): there is so  so much code


Looking forward to it!

niklas88 · 2018-08-11T19:33:31Z

src/index/IndexMetaData.h

-
-  bool hasBlocks() const { return _rmdPairs.hasBlocks(); }
+  // persistentRMD == true means we do not need to add rmd to _data
+  // but assume that it is already is contained in _data (for persistent


one "is" too many

niklas88 · 2018-08-11T19:33:45Z

src/index/IndexMetaData.h


-  uint8_t getCol2LogMultiplicity() const {
-    return _rmdPairs.getCol2LogMultiplicity();
+  // Persistent (called MmapBased here) have to be separated from RAM-based


word missing

niklas88 · 2018-08-11T19:36:43Z

src/index/IndexMetaDataImpl.h

+    const auto el = *it;
+    // when we are MmapBased, the _data member is already persistent on disk, so
+    // we do not write it here
+    if constexpr (!imd._isMmapBased) {


…tations - Almost every token in the vocabulary occurs as an object in a triple, so the OPS and OSP MetaData are Dense (all possible keys/vocabIds occur). This makes using array-like data structures better. - Implemented a class MmapVector that is able to handle a persistent dynamic array that is stored on hard disk and made accessible from RAM via mmap - Implemented a class MmapVectorView that is a readonly version of MmapVector - Templated the IndexMetaDataClass to handle either the "traditional" HashMap based IndexMetaData or the new Mmap based ones. - Implemented a converter to transform previously built indices to the new format - Currently, the OSP and OPS permutations are always mmap based whereas the rest is hashMap based. One could also change SOP and SPO to mmap based to save even more Ram or implement a dynamic dispatching at runtime. - Updated README.md with documentation on the MetaDataConverter and information about the reduced RAM usage for OSP and OPS

niklas88 reviewed Aug 1, 2018

View reviewed changes

niklas88 reviewed Aug 3, 2018

View reviewed changes

src/util/MmapVector.h Outdated

size_t _bytesize;

};

// 2 tags two differentiate different versions of the

Copy link

Member

niklas88 Aug 3, 2018

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

s/two/to/

niklas88 reviewed Aug 3, 2018

View reviewed changes

niklas88 reviewed Aug 6, 2018

View reviewed changes

niklas88 reviewed Aug 7, 2018

View reviewed changes

joka921 force-pushed the f.mmapBasedArray branch from e2b9597 to 7a7ac0a Compare August 10, 2018 13:50

joka921 force-pushed the f.mmapBasedArray branch from 678c9e3 to 1a2398e Compare August 10, 2018 17:17

joka921 force-pushed the f.mmapBasedArray branch from c30628b to 2da241d Compare August 11, 2018 14:03

niklas88 approved these changes Aug 11, 2018

View reviewed changes

joka921 force-pushed the f.mmapBasedArray branch from 2da241d to 5194941 Compare August 12, 2018 07:47

joka921 merged commit 7b6e89b into ad-freiburg:master Aug 12, 2018

joka921 deleted the f.mmapBasedArray branch August 23, 2022 20:40

Dense Meta Data using MMap-Based arrays #80

Dense Meta Data using MMap-Based arrays #80

Conversation

joka921 commented Jul 31, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

niklas88 left a comment • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

niklas88 commented Aug 3, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

niklas88 commented Aug 6, 2018

joka921 commented Aug 6, 2018

niklas88 commented Aug 6, 2018

niklas88 commented Aug 6, 2018

niklas88 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

niklas88 commented Aug 7, 2018 • edited

niklas88 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

niklas88 commented Aug 10, 2018

joka921 commented Aug 10, 2018

niklas88 commented Aug 11, 2018

joka921 commented Aug 11, 2018

niklas88 left a comment

Choose a reason for hiding this comment

niklas88 Aug 11, 2018 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

niklas88 left a comment •

edited

niklas88 commented Aug 7, 2018 •

edited

niklas88 Aug 11, 2018 •

edited