A library for creating perfect hash tables.
This README is a work-in-progress.
Usage: PerfectHashBulkCreate.exe
<KeysDirectory> <OutputDirectory>
<Algorithm> <HashFunction> <MaskFunction>
<MaximumConcurrency>
[BulkCreateFlags] [KeysLoadFlags] [TableCreateFlags]
[TableCompileFlags] [TableCreateParameters]
Bulk Create Flags:
--SkipTestAfterCreate
Normally, after a table has been successfully created, it is tested.
Setting this flag disables this behavior.
N.B. This will also disable benchmarking, so no performance information
will be present in the .csv output file.
--Compile
Compiles the table after creation.
N.B. msbuild.exe must be on the PATH environment variable for this
to work. (The current error message if this isn't the case is
quite cryptic.)
Keys Load Flags:
--TryLargePagesForKeysData
Tries to allocate the keys buffer using large pages.
--SkipKeysVerification
Skips the logic that enumerates all keys after loading and a) ensures
they are sorted, and b) constructs a keys bitmap. If you can be certain
the keys are sorted, specifying this flag can provide a speedup when
loading large key sets.
--DisableImplicitKeyDownsizing
When loading keys that are 64-bit (8 bytes), a bitmap is kept that
tracks whether or not a given bit was seen across the entire key set.
After enumerating the set, the number of zeros in the bitmap are
counted; if this number is less than or equal to 32, it means that the
entire key set can be compressed into 32-bit values with some parallel
bit extraction logic (i.e. _pext_u64()). As this has beneficial size
and performance implications, when detected, the key load operation will
implicitly heap-allocate another array and convert all the 64-bit keys
into their unique 32-bit equivalent. Specifying this flag will disable
this behavior.
--TryInferKeySizeFromKeysFilename
The default key size is 32-bit (4 bytes). When this flag is present,
if the keys file name ends with "64.keys" (e.g. "foo64.keys"), the key
size will be interpreted as 64-bit (8 bytes). This flag takes
precedence over the table create parameter --KeySizeInBytes.
Table Create Flags:
--Silent
Disables console printing of the dots, dashes and other characters used
to (crudely) visualize the result of individual table create operations.
--NoFileIo
Disables writing of all files when a perfect hash solution has been
found. The only time you would use this flag from the console
application is to observe the performance of table creation without
performing any file I/O operations.
--Paranoid
Enables redundant checks in the routine that determines whether or not
a generated graph is acyclic.
--FirstGraphWins | --FindBestGraph
--FirstGraphWins [default]
This is the default behavior. When searching for solutions in
parallel, the first graph to be found, "wins". i.e. it's the
solution that is subsequently written to disk.
--FindBestGraph
Requires the following two table create parameters to be present:
--BestCoverageAttempts=N
--BestCoverageType=<CoverageType>
The table create routine will then run until it finds the number of
best coverage attempts specified. At that point, the graph that was
found to be the "best" based on the coverage type predicate "wins",
and is subsequently saved to disk.
N.B. This option is significantly more CPU intensive than the
--FirstGraphWins mode, however, it is highly probable that it
will find a graph that is better (based on the predicate) than
when in first graph wins mode.
--SkipMemoryCoverageInFirstGraphWinsMode
Skips calculating memory coverage information when in "first graph wins"
mode. This will result in the corresponding fields in the .csv output
indicating 0.
--SkipGraphVerification
When present, skips the internal graph verification check that ensures
a valid perfect hash solution has been found (i.e. with no collisions
across the entire key set).
--DisableCsvOutputFile
When present, disables writing the .csv output file. This is required
when running multiple instances of the tool against the same output
directory in parallel.
--OmitCsvRowIfTableCreateFailed
When present, omits writing a row in the .csv output file if table
creation fails for a given keys file. Ignored if --DisableCsvOutputFile
is speficied.
--OmitCsvRowIfTableCreateSucceeded
When present, omits writing a row in the .csv output file if table
creation succeeded for a given keys file. Ignored if
--DisableCsvOutputFile is specified.
--IndexOnly
When set, affects the generated C files by defining the C preprocessor
macro CPH_INDEX_ONLY, which results in omitting the compiled perfect
hash routines that deal with the underlying table values array (i.e.
any routine other than Index(); e.g. Insert(), Lookup(), Delete() etc),
as well as the array itself. This results in a size reduction of the
final compiled perfect hash binary. Additionally, only the .dll and
BenchmarkIndex projects will be built, as the BenchmarkFull and Test
projects require access to a table values array. This flag is intended
to be used if you only need the Index() routine and will be managing the
table values array independently.
--UseRwsSectionForTableValues
When set, tells the linker to use a shared read-write section for the
table values array, e.g.: #pragma comment(linker,"/section:.cphval,rws")
This will result in the table values array being accessible across
multiple processes. Thus, the array will persist as long as one process
maintains an open section (mapping); i.e. keeps the .dll loaded.
--UseNonTemporalAvx2Routines
When set, uses implementations of RtlCopyPages and RtlFillPages that
use non-temporal hints.
--UsePreviousTableSize
When set, uses any previously-recorded table sizes associated with
the keys file for the given algorithm, hash function and masking type.
N.B. To forcibly delete all previously-recorded table sizes from all
keys in a directory, the following PowerShell snippet can be used:
PS C:\Temp\keys> Get-Item -Path *.keys -Stream *.TableSize | Remove-Item
--IncludeNumberOfTableResizeEventsInOutputPath
When set, incorporates the number of table resize events encountered
whilst searching for a perfect hash solution into the final output
names, e.g.:
C:\Temp\output\KernelBase_2485_1_Chm01_Crc32Rotate_And\...
^
Number of resize events.
--IncludeNumberOfTableElementsInOutputPath
When set, incorporates the number of table elements (i.e. the final
table size) into the output path, e.g.:
C:\Temp\output\KernelBase_2485_16384_Chm01_Crc32Rotate_And\...
^
Number of table elements.
N.B. These two flags can be combined, yielding a path as follows:
C:\Temp\output\KernelBase_2485_1_16384_Chm01_Crc32Rotate_And\...
N.B. It is important to understand how table resize events impact the
behavior of this program if one or both of these flags are present.
Using the example above, the initial path that will be used for
the solution will be:
C:\Temp\output\KernelBase_2485_0_8192_Chm01_Crc32Rotate_And\...
After the maximum number of attempts are reached, a table resize
event occurs; the new path component will be:
C:\Temp\output\KernelBase_2485_1_16384_Chm01_Crc32Rotate_And\...
However, the actual renaming of the directory is not done until
a solution has been found and all files have been written. If
this program is being run repeatedly, then the target directory
will already exist. This complicates things, as, unlike files,
we can't just replace an existing directory with a new one.
There are two ways this could be handled: a) delete all the
existing files under the target path, then delete the directory,
then perform the rename, or b) move the target directory somewhere
else first, preserving the existing contents, then proceed with
the rename.
This program takes the latter approach. The existing directory
will be moved to:
C:\Temp\output\old\KernelBase_1_16384_Chm01_Crc32Rotate_And_2018-11-19-011023-512\...
The timestamp appended to the directory name is derived from the
existing directory's creation time, which should ensure uniqueness.
(In the unlikely situation the target directory already exists in
the old subdirectory, the whole operation is aborted and the table
create routine returns a failure.)
The point of mentioning all of this is the following: when one or
both of these flags are routinely specified, the number of output
files rooted in the output directory's 'old' subdirectory will grow
rapidly, consuming a lot of disk space. Thus, if the old files are
not required, it is recommended to regularly delete them manually.
Table Compile Flags:
N/A
Table Create Parameters:
--ValueSizeInBytes=4|8
Sets the size, in bytes, of the value element that will be stored in the
compiled perfect hash table via Insert(). Defaults to 4 bytes (ULONG).
--MainWorkThreadpoolPriority=<High|Normal|Low> [default: Normal]
--FileWorkThreadpoolPriority=<High|Normal|Low> [default: Normal]
Sets the main work (i.e. the CPU-intensive graph solving) threadpool
priority, or the file work threadpool priority, to the given value.
--AttemptsBeforeTableResize=N [default = 100,000,000,000 ]
Specifies the number of attempts at solving the graph that will be made
before a table resize event will occur (assuming that resize events are
permitted, as per the following flag).
--MaxNumberOfTableResizes=N [default = 5]
Maximum number of table resizes that will be permitted before giving up.
--BestCoverageAttempts=N
Where N is a positive integer, and represents the number of attempts
that will be made at finding a "best" graph (based on the best coverage
type requested below) before the create table routine returns.
--BestCoverageType=<CoverageType>
Indicates the predicate to determine what constitutes the best graph.
Valid coverage types:
HighestNumberOfEmptyPages
HighestNumberOfEmptyLargePages
HighestNumberOfEmptyCacheLines
HighestMaxGraphTraversalDepth
HighestTotalGraphTraversals
HighestMaxAssignedPerCacheLineCount
LowestNumberOfEmptyPages
LowestNumberOfEmptyLargePages
LowestNumberOfEmptyCacheLines
LowestMaxGraphTraversalDepth
LowestTotalGraphTraversals
LowestMaxAssignedPerCacheLineCount
The following predicates must be used in conjunction with --KeysSubset:
HighestMaxAssignedPerCacheLineCountForKeysSubset
HighestNumberOfPagesUsedByKeysSubset
HighestNumberOfLargePagesUsedByKeysSubset
HighestNumberOfCacheLinesUsedByKeysSubset
LowestMaxAssignedPerCacheLineCountForKeysSubset
LowestNumberOfPagesUsedByKeysSubset
LowestNumberOfLargePagesUsedByKeysSubset
LowestNumberOfCacheLinesUsedByKeysSubset
Console Output Character Legend
Char | Meaning
. Table created successfully.
+ Table resize event occured.
x Failed to create a table. The maximum number of attempts at trying to
solve the table at a given size was reached, and no more resize attempts
were possible (due to the maximum resize limit also being hit).
F Failed to create a table due to a target not being reached by a specific
number of attempts or time duration. Not yet implemented.
* None of the worker threads were able to allocate sufficient memory to
attempt solving the graph.
! The system is out of memory.
L The system is running low on memory (a low memory event is triggered
at about 90%% RAM usage). In certain situations we can detect this
situation prior to actually running out of memory; in these cases,
we abort the current table creation attempt (which will instantly
relieve system memory pressure).
V The graph was created successfully, however, we weren't able to allocate
enough memory for the table values array in order for the array to be
used after creation. This can be avoided by omitting --TestAfterCreate.
T The requested number of table elements was too large (exceeded 32 bits).
S A shutdown event was received. This shouldn't be seen unless externally
signaling the named shutdown event associated with a context.
Algorithms:
ID | Name
1 Chm01
Hash Functions:
ID | Name (Number of Seeds)
2 Jenkins (2)
12 Fnv (2)
14 Crc32RotateX (3)
15 Crc32RotateXY (3)
16 Crc32RotateWXYZ (3)
17 RotateMultiplyXorRotate (3)
18 ShiftMultiplyXorShift (3)
19 ShiftMultiplyXorShift2 (6)
20 RotateMultiplyXorRotate2 (6)
21 MultiplyRotateR (3)
22 MultiplyRotateLR (3)
23 MultiplyShiftR (3)
24 MultiplyShiftLR (3)
25 Multiply (2)
26 MultiplyXor (4)
Mask Functions:
ID | Name
1 Modulus (does not work!)
2 And
Creates a single perfect hash table. Provides the ability to specify a subset of keys as part of the "find best graph" predicates.
Usage: PerfectHashCreate.exe
<KeysPath> <OutputDirectory>
<Algorithm> <HashFunction> <MaskFunction>
<MaximumConcurrency>
[KeysLoadFlags] [TableCreateFlags]
[TableCompileFlags] [TableCreateParameters]
Keys Load Flags:
--TryLargePagesForKeysData
Tries to allocate the keys buffer using large pages.
--SkipKeysVerification
Skips the logic that enumerates all keys after loading and a) ensures
they are sorted, and b) constructs a keys bitmap. If you can be certain
the keys are sorted, specifying this flag can provide a speedup when
loading large key sets.
--DisableImplicitKeyDownsizing
When loading keys that are 64-bit (8 bytes), a bitmap is kept that
tracks whether or not a given bit was seen across the entire key set.
After enumerating the set, the number of zeros in the bitmap are
counted; if this number is less than or equal to 32, it means that the
entire key set can be compressed into 32-bit values with some parallel
bit extraction logic (i.e. _pext_u64()). As this has beneficial size
and performance implications, when detected, the key load operation will
implicitly heap-allocate another array and convert all the 64-bit keys
into their unique 32-bit equivalent. Specifying this flag will disable
this behavior.
--TryInferKeySizeFromKeysFilename
The default key size is 32-bit (4 bytes). When this flag is present,
if the keys file name ends with "64.keys" (e.g. "foo64.keys"), the key
size will be interpreted as 64-bit (8 bytes). This flag takes
precedence over the table create parameter --KeySizeInBytes.
Table Create Flags:
--SkipTestAfterCreate
Normally, after a table has been successfully created, it is tested.
Setting this flag disables this behavior.
N.B. This will also disable benchmarking, so no performance information
will be present in the .csv output file.
--Silent
Disables console printing of the dots, dashes and other characters used
to (crudely) visualize the result of individual table create operations.
--NoFileIo
Disables writing of all files when a perfect hash solution has been
found. The only time you would use this flag from the console
application is to observe the performance of table creation without
performing any file I/O operations.
--Paranoid
Enables redundant checks in the routine that determines whether or not
a generated graph is acyclic.
--FirstGraphWins | --FindBestGraph
--FirstGraphWins [default]
This is the default behavior. When searching for solutions in
parallel, the first graph to be found, "wins". i.e. it's the
solution that is subsequently written to disk.
--FindBestGraph
Requires the following two table create parameters to be present:
--BestCoverageAttempts=N
--BestCoverageType=<CoverageType>
The table create routine will then run until it finds the number of
best coverage attempts specified. At that point, the graph that was
found to be the "best" based on the coverage type predicate "wins",
and is subsequently saved to disk.
N.B. This option is significantly more CPU intensive than the
--FirstGraphWins mode, however, it is highly probable that it
will find a graph that is better (based on the predicate) than
when in first graph wins mode.
--SkipMemoryCoverageInFirstGraphWinsMode
Skips calculating memory coverage information when in "first graph wins"
mode. This will result in the corresponding fields in the .csv output
indicating 0.
--SkipGraphVerification
When present, skips the internal graph verification check that ensures
a valid perfect hash solution has been found (i.e. with no collisions
across the entire key set).
--DisableCsvOutputFile
When present, disables writing the .csv output file. This is required
when running multiple instances of the tool against the same output
directory in parallel.
--OmitCsvRowIfTableCreateFailed
When present, omits writing a row in the .csv output file if table
creation fails for a given keys file. Ignored if --DisableCsvOutputFile
is speficied.
--OmitCsvRowIfTableCreateSucceeded
When present, omits writing a row in the .csv output file if table
creation succeeded for a given keys file. Ignored if
--DisableCsvOutputFile is specified.
--IndexOnly
When set, affects the generated C files by defining the C preprocessor
macro CPH_INDEX_ONLY, which results in omitting the compiled perfect
hash routines that deal with the underlying table values array (i.e.
any routine other than Index(); e.g. Insert(), Lookup(), Delete() etc),
as well as the array itself. This results in a size reduction of the
final compiled perfect hash binary. Additionally, only the .dll and
BenchmarkIndex projects will be built, as the BenchmarkFull and Test
projects require access to a table values array. This flag is intended
to be used if you only need the Index() routine and will be managing the
table values array independently.
--UseRwsSectionForTableValues
When set, tells the linker to use a shared read-write section for the
table values array, e.g.: #pragma comment(linker,"/section:.cphval,rws")
This will result in the table values array being accessible across
multiple processes. Thus, the array will persist as long as one process
maintains an open section (mapping); i.e. keeps the .dll loaded.
--UseNonTemporalAvx2Routines
When set, uses implementations of RtlCopyPages and RtlFillPages that
use non-temporal hints.
--UsePreviousTableSize
When set, uses any previously-recorded table sizes associated with
the keys file for the given algorithm, hash function and masking type.
N.B. To forcibly delete all previously-recorded table sizes from all
keys in a directory, the following PowerShell snippet can be used:
PS C:\Temp\keys> Get-Item -Path *.keys -Stream *.TableSize | Remove-Item
--IncludeNumberOfTableResizeEventsInOutputPath
When set, incorporates the number of table resize events encountered
whilst searching for a perfect hash solution into the final output
names, e.g.:
C:\Temp\output\KernelBase_2485_1_Chm01_Crc32Rotate_And\...
^
Number of resize events.
--IncludeNumberOfTableElementsInOutputPath
When set, incorporates the number of table elements (i.e. the final
table size) into the output path, e.g.:
C:\Temp\output\KernelBase_2485_16384_Chm01_Crc32Rotate_And\...
^
Number of table elements.
N.B. These two flags can be combined, yielding a path as follows:
C:\Temp\output\KernelBase_2485_1_16384_Chm01_Crc32Rotate_And\...
N.B. It is important to understand how table resize events impact the
behavior of this program if one or both of these flags are present.
Using the example above, the initial path that will be used for
the solution will be:
C:\Temp\output\KernelBase_2485_0_8192_Chm01_Crc32Rotate_And\...
After the maximum number of attempts are reached, a table resize
event occurs; the new path component will be:
C:\Temp\output\KernelBase_2485_1_16384_Chm01_Crc32Rotate_And\...
However, the actual renaming of the directory is not done until
a solution has been found and all files have been written. If
this program is being run repeatedly, then the target directory
will already exist. This complicates things, as, unlike files,
we can't just replace an existing directory with a new one.
There are two ways this could be handled: a) delete all the
existing files under the target path, then delete the directory,
then perform the rename, or b) move the target directory somewhere
else first, preserving the existing contents, then proceed with
the rename.
This program takes the latter approach. The existing directory
will be moved to:
C:\Temp\output\old\KernelBase_1_16384_Chm01_Crc32Rotate_And_2018-11-19-011023-512\...
The timestamp appended to the directory name is derived from the
existing directory's creation time, which should ensure uniqueness.
(In the unlikely situation the target directory already exists in
the old subdirectory, the whole operation is aborted and the table
create routine returns a failure.)
The point of mentioning all of this is the following: when one or
both of these flags are routinely specified, the number of output
files rooted in the output directory's 'old' subdirectory will grow
rapidly, consuming a lot of disk space. Thus, if the old files are
not required, it is recommended to regularly delete them manually.
Table Compile Flags:
N/A
Table Create Parameters:
--ValueSizeInBytes=4|8
Sets the size, in bytes, of the value element that will be stored in the
compiled perfect hash table via Insert(). Defaults to 4 bytes (ULONG).
--MainWorkThreadpoolPriority=<High|Normal|Low> [default: Normal]
--FileWorkThreadpoolPriority=<High|Normal|Low> [default: Normal]
Sets the main work (i.e. the CPU-intensive graph solving) threadpool
priority, or the file work threadpool priority, to the given value.
--AttemptsBeforeTableResize=N [default = 100,000,000,000 ]
Specifies the number of attempts at solving the graph that will be made
before a table resize event will occur (assuming that resize events are
permitted, as per the following flag).
--MaxNumberOfTableResizes=N [default = 5]
Maximum number of table resizes that will be permitted before giving up.
--BestCoverageAttempts=N
Where N is a positive integer, and represents the number of attempts
that will be made at finding a "best" graph (based on the best coverage
type requested below) before the create table routine returns.
--BestCoverageType=<CoverageType>
Indicates the predicate to determine what constitutes the best graph.
Valid coverage types:
HighestNumberOfEmptyPages
HighestNumberOfEmptyLargePages
HighestNumberOfEmptyCacheLines
HighestMaxGraphTraversalDepth
HighestTotalGraphTraversals
HighestMaxAssignedPerCacheLineCount
LowestNumberOfEmptyPages
LowestNumberOfEmptyLargePages
LowestNumberOfEmptyCacheLines
LowestMaxGraphTraversalDepth
LowestTotalGraphTraversals
LowestMaxAssignedPerCacheLineCount
The following predicates must be used in conjunction with --KeysSubset:
HighestMaxAssignedPerCacheLineCountForKeysSubset
HighestNumberOfPagesUsedByKeysSubset
HighestNumberOfLargePagesUsedByKeysSubset
HighestNumberOfCacheLinesUsedByKeysSubset
LowestMaxAssignedPerCacheLineCountForKeysSubset
LowestNumberOfPagesUsedByKeysSubset
LowestNumberOfLargePagesUsedByKeysSubset
LowestNumberOfCacheLinesUsedByKeysSubset
--KeysSubset=N,N+1[,N+2,N+3,...] (e.g. --KeysSubset=10,50,123,600,670)
Supplies a comma-separated list of keys in ascending key-value order.
Must contain two or more elements.
Algorithms:
ID | Name
1 Chm01
Hash Functions:
ID | Name (Number of Seeds)
2 Jenkins (2)
12 Fnv (2)
14 Crc32RotateX (3)
15 Crc32RotateXY (3)
16 Crc32RotateWXYZ (3)
17 RotateMultiplyXorRotate (3)
18 ShiftMultiplyXorShift (3)
19 ShiftMultiplyXorShift2 (6)
20 RotateMultiplyXorRotate2 (6)
21 MultiplyRotateR (3)
22 MultiplyRotateLR (3)
23 MultiplyShiftR (3)
24 MultiplyShiftLR (3)
25 Multiply (2)
26 MultiplyXor (4)
Mask Functions:
ID | Name
1 Modulus (does not work!)
2 And
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent