Summary:
I was investigating performance issues in the SstFileWriter and found all of the following:
- The SstFileWriter::Add() function created a local InternalKey every time it was called generating a allocation and free each time. Changed to have an InternalKey member variable that can be reset with the new InternalKey::Set() function.
- In SstFileWriter::Add() the smallest_key and largest_key values were assigned the result of a ToString() call, but it is simpler to just assign them directly from the user's key.
- The Slice class had no move constructor so each time one was returned from a function a new one had to be allocated, the old data copied to the new, and the old one was freed. I added the move constructor which also required a copy constructor and assignment operator.
- The BlockBuilder::CurrentSizeEstimate() function calculates the current estimate size, but was being called 2 or 3 times for each key added. I changed the class to maintain a running estimate (equal to the original calculation) so that the function can return an already calculated value.
- The code in BlockBuilder::Add() that calculated the shared bytes between the last key and the new key duplicated what Slice::difference_offset does, so I replaced it with the standard function.
- BlockBuilder::Add() had code to copy just the changed portion into the last key value (and asserted that it now matched the new key). It is more efficient just to copy the whole new key over.
- Moved this same code up into the 'if (use_delta_encoding_)' since the last key value is only needed when delta encoding is on.
- FlushBlockBySizePolicy::BlockAlmostFull calculated a standard deviation value each time it was called, but this information would only change if block_size of block_size_deviation changed, so I created a member variable to hold the value to avoid the calculation each time.
- Each PutVarint??() function has a buffer and calls std::string::append(). Two or three calls in a row could share a buffer and a single call to std::string::append().
Some of these will be helpful outside of the SstFileWriter. I'm not 100% the addition of the move constructor is appropriate as I wonder why this wasn't done before - maybe because of compiler compatibility? I tried it on gcc 4.8 and 4.9.
Test Plan: The changes should not affect the results so the existing tests should all still work and no new tests were added. The value of the changes was seen by manually testing the SstFileWriter class through MyRocks and adding timing code to identify problem areas.
Reviewers: sdong, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D59607
Summary: Add option write_buffer_manager to help users control total memory spent on memtables across multiple DB instances.
Test Plan: Add a new unit test.
Reviewers: yhchiang, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: adela, benj, sumeet, muthu, leveldb, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D59925
Summary: SstFileWriter ignore Options::bottommost_compression, update it to use bottommost_compression if available
Test Plan:
make check -j64
verified used compression using ./sst_dump
Reviewers: sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba, yoshinorim
Differential Revision: https://reviews.facebook.net/D59841
Summary:
Right now the way we do BlockIter::Prev() is like this
- Go to the beginning of the restart interval
- Keep moving forward (and decoding keys using ParseNextKey()) until we reach the desired key
This can be optimized by caching the decoded entries in the first pass and reusing them in consecutive BlockIter::Prev() calls
Before caching
```
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readreverse" --db="/dev/shm/bench_prev_opt/" --use_existing_db --disable_auto_compactions
DB path: [/dev/shm/bench_prev_opt/]
readreverse : 0.413 micros/op 2423972 ops/sec; 268.2 MB/s
DB path: [/dev/shm/bench_prev_opt/]
readreverse : 0.414 micros/op 2413867 ops/sec; 267.0 MB/s
DB path: [/dev/shm/bench_prev_opt/]
readreverse : 0.410 micros/op 2440881 ops/sec; 270.0 MB/s
DB path: [/dev/shm/bench_prev_opt/]
readreverse : 0.414 micros/op 2417298 ops/sec; 267.4 MB/s
DB path: [/dev/shm/bench_prev_opt/]
readreverse : 0.413 micros/op 2421682 ops/sec; 267.9 MB/s
```
After caching
```
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readreverse" --db="/dev/shm/bench_prev_opt/" --use_existing_db --disable_auto_compactions
DB path: [/dev/shm/bench_prev_opt/]
readreverse : 0.324 micros/op 3088955 ops/sec; 341.7 MB/s
DB path: [/dev/shm/bench_prev_opt/]
readreverse : 0.335 micros/op 2980999 ops/sec; 329.8 MB/s
DB path: [/dev/shm/bench_prev_opt/]
readreverse : 0.341 micros/op 2929681 ops/sec; 324.1 MB/s
DB path: [/dev/shm/bench_prev_opt/]
readreverse : 0.344 micros/op 2908490 ops/sec; 321.8 MB/s
DB path: [/dev/shm/bench_prev_opt/]
readreverse : 0.338 micros/op 2958404 ops/sec; 327.3 MB/s
```
Test Plan: COMPILE_WITH_ASAN=1 make check -j64
Reviewers: andrewkr, yiwu, sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba, yoshinorim
Differential Revision: https://reviews.facebook.net/D59463
Summary:
We have alot of code duplication whenever we call FullMerge we keep duplicating the instrumentation and statistics code
This is a simple diff to refactor the code to use TimedFullMerge instead of FullMerge
Test Plan: COMPILE_WITH_ASAN=1 make check -j64
Reviewers: andrewkr, yhchiang, sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D59577
Summary:
Try to decompress compressed blocks when a special flag is set.
assert and crash in debug builds if we can't decompress the just-compressed input.
Test Plan: Run unit-tests.
Reviewers: dhruba, andrewkr, sdong, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D59145
Summary: Currently, if users define both of full key bloom and prefix bloom in SST files. During Get(), if full key bloom shows the key may exist, we still go ahead and check prefix bloom. This is wasteful. If bloom filter for full keys exists, we should always ignore prefix bloom in Get().
Test Plan: Run existing tests
Reviewers: yhchiang, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: leveldb, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D57825
Summary: With `table_options.cache_index_and_filter_blocks = true`, index and filter blocks are stored in block cache. Then people are curious how much of the block cache total size is used by indexes and bloom filters. It will be nice we have a way to report that. It can help people tune performance and plan for optimized hardware setting. We add several enum values for db Statistics. BLOCK_CACHE_INDEX/FILTER_BYTES_INSERT - BLOCK_CACHE_INDEX/FILTER_BYTES_ERASE = current INDEX/FILTER total block size in bytes.
Test Plan:
write a test case called `DBBlockCacheTest.IndexAndFilterBlocksStats`. The result is:
```
[gzh@dev9927.prn1 ~/local/rocksdb] make db_block_cache_test -j64 && ./db_block_cache_test --gtest_filter=DBBlockCacheTest.IndexAndFilterBlocksStats
Makefile:101: Warning: Compiling in debug mode. Don't use the resulting binary in production
GEN util/build_version.cc
make: `db_block_cache_test' is up to date.
Note: Google Test filter = DBBlockCacheTest.IndexAndFilterBlocksStats
[==========] Running 1 test from 1 test case.
[----------] Global test environment set-up.
[----------] 1 test from DBBlockCacheTest
[ RUN ] DBBlockCacheTest.IndexAndFilterBlocksStats
[ OK ] DBBlockCacheTest.IndexAndFilterBlocksStats (689 ms)
[----------] 1 test from DBBlockCacheTest (689 ms total)
[----------] Global test environment tear-down
[==========] 1 test from 1 test case ran. (689 ms total)
[ PASSED ] 1 test.
```
Reviewers: IslamAbdelRahman, andrewkr, sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D58677
Summary: Deprecate this one option and delete code and tests that are now superfluous.
Test Plan: all tests pass
Reviewers: igor, yhchiang, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: msalib, leveldb, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D55317
Summary:
Added a new abstraction to cache page to RocksDB designed for the read
cache use.
RocksDB current block cache is more of an object cache. For the persistent read cache
project, what we need is a page cache equivalent. This changes adds a cache
abstraction to RocksDB to cache pages called PersistentCache. PersistentCache can cache
uncompressed pages or raw pages (content as in filesystem). The user can
choose to operate PersistentCache either in COMPRESSED or UNCOMPRESSED mode.
Blame Rev:
Test Plan: Run unit tests
Reviewers: sdong
Subscribers: andrewkr, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D55707
Summary: Store SST file compression algorithm as a TableProperty.
Test Plan: Modified and ran the table_test UT that checks for TableProperties
Reviewers: IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: lgalanis, andrewkr, dhruba, IslamAbdelRahman
Differential Revision: https://reviews.facebook.net/D58017
Summary: This is to provide a way for users to skip prefix bloom in point look-up.
Test Plan: Add a new unit test scenario.
Reviewers: IslamAbdelRahman
Subscribers: leveldb, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D57747
Summary:
This adds a new metablock containing a shared dictionary that is used
to compress all data blocks in the SST file. The size of the shared dictionary
is configurable in CompressionOptions and defaults to 0. It's currently only
used for zlib/lz4/lz4hc, but the block will be stored in the SST regardless of
the compression type if the user chooses a nonzero dictionary size.
During compaction, computes the dictionary by randomly sampling the first
output file in each subcompaction. It pre-computes the intervals to sample
by assuming the output file will have the maximum allowable length. In case
the file is smaller, some of the pre-computed sampling intervals can be beyond
end-of-file, in which case we skip over those samples and the dictionary will
be a bit smaller. After the dictionary is generated using the first file in a
subcompaction, it is loaded into the compression library before writing each
block in each subsequent file of that subcompaction.
On the read path, gets the dictionary from the metablock, if it exists. Then,
loads that dictionary into the compression library before reading each block.
Test Plan: new unit test
Reviewers: yhchiang, IslamAbdelRahman, cyan, sdong
Reviewed By: sdong
Subscribers: andrewkr, yoshinorim, kradhakrishnan, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D52287
Summary:
While trying to reuse PinData() / ReleasePinnedData() .. to optimize away some memcpys I realized that there is a significant overhead for using PinData() / ReleasePinnedData if they were called many times.
This diff refactor the pinning logic by introducing PinnedIteratorsManager a centralized component that will be created once and will be notified whenever we need to Pin an Iterator. This implementation have much less overhead than the original implementation
Test Plan:
make check -j64
COMPILE_WITH_ASAN=1 make check -j64
Reviewers: yhchiang, sdong, andrewkr
Reviewed By: andrewkr
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D56493
Summary:
This is the original diff that I have landed and reverted and now I want to land again https://reviews.facebook.net/D34269
For old SST files we will show
```
comparator name: N/A
merge operator name: N/A
property collectors names: N/A
```
For new SST files with no merge operator name and with no property collectors
```
comparator name: leveldb.BytewiseComparator
merge operator name: nullptr
property collectors names: []
```
for new SST files with these properties
```
comparator name: leveldb.BytewiseComparator
merge operator name: UInt64AddOperator
property collectors names: [DummyPropertiesCollector1,DummyPropertiesCollector2]
```
Test Plan: unittests
Reviewers: andrewkr, yhchiang, sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D56487
Comparable with Snappy on comp ratio.
Implemented using Windows API, does not require external package.
Avaiable since Windows 8 and server 2012.
Use -DXPRESS=1 with CMake to enable.
Summary:
In the case where we can't find a filter block, there is not much benefit of doing the binary search and see whether the index key has the prefix. With the change, we blindly return true if we can't get the filter.
It also fixes missing row cases for reverse comparator with full bloom.
Test Plan: Add a test case that used to fail.
Reviewers: yhchiang, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: kradhakrishnan, yiwu, hermanlee4, yoshinorim, leveldb, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D56697
Summary:
Full block checking should be a good enough indication of prefix existance. No need to further check data block.
This also fixes wrong results when using prefix bloom and reverse bitwise comparator.
Test Plan: Will add a unit test.
Reviewers: yhchiang, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: hermanlee4, yoshinorim, yiwu, kradhakrishnan, leveldb, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D56625
Summary: PrefixTest.PrefixAndWholeKeyTest runs against the same directory as prefix_test, which sometimes fail parallel tests. Fix it.
Test Plan: Run it in parallel and see it doesn't fail anymore.
Reviewers: andrewkr
Reviewed By: andrewkr
Subscribers: leveldb, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D56541
Summary: Cache shard bit 4 is sometimes too small and 6 is a more common value picked by users. Make that default. It shouldn't hurt much to change options.max_file_opening_threads default to be 16, which will reduce the worst case DB open time.
Test Plan: Run all existing tests.
Reviewers: IslamAbdelRahman, yhchiang, kradhakrishnan, igor, andrewkr
Reviewed By: andrewkr
Subscribers: andrewkr, MarkCallaghan, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D55047
Summary:
Added the column family name to the properties block. This property
is omitted only if the property is unavailable, such as when RepairDB()
writes SST files.
In a next diff, I will change RepairDB to use this new property for
deciding to which column family an existing SST file belongs. If this
property is missing, it will add it to the "unknown" column family (same
as its existing behavior).
Test Plan:
New unit test:
$ ./db_table_properties_test --gtest_filter=DBTablePropertiesTest.GetColumnFamilyNameProperty
Reviewers: IslamAbdelRahman, yhchiang, sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D55605
Summary:
When a block based table file is opened, if prefetch_index_and_filter is true, it will prefetch the index and filter blocks, putting them into the block cache.
What this feature adds: when a L0 block based table file is opened, if pin_l0_filter_and_index_blocks_in_cache is true in the options (and prefetch_index_and_filter is true), then the filter and index blocks aren't released back to the block cache at the end of BlockBasedTableReader::Open(). Instead the table reader takes ownership of them, hence pinning them, ie. the LRU cache will never push them out. Meanwhile in the table reader, further accesses will not hit the block cache, thus avoiding lock contention.
Test Plan:
'export TEST_TMPDIR=/dev/shm/ && DISABLE_JEMALLOC=1 OPT=-g make all valgrind_check -j32' is OK.
I didn't run the Java tests, I don't have Java set up on my devserver.
Reviewers: sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D56133
Summary:
Expose the inverse of ToString(hex=true) on Slice: Slice::DecodeHex
Refactor the other implementation of to/from hex in ldb_cmd.h to use the Slice
version
(Difference between the 2 is whether 0x is expected/produced in front of the hex
string or not)
Eliminated support for invalid odd length hex string - this is now invalid
instead of having 1/2 byte set
Added (inverse of HexToString) test for LDBCommand::StringToHex which also
indirectly tests Slice::ToString(true)
After moving the original implementation from ldb_cmd.h, updated it to much simpler/efficient version
(originally/inspired from https://github.com/facebook/wdt/blob/master/util/EncryptionUtils.cpp#L140-L169 )
Test Plan: run tests
Reviewers: uddipta, sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D56121
Summary:
When a block based table file is opened, if prefetch_index_and_filter is true, it will prefetch the index and filter blocks, putting them into the block cache.
What this feature adds: when a L0 block based table file is opened, if pin_l0_filter_and_index_blocks_in_cache is true in the options (and prefetch_index_and_filter is true), then the filter and index blocks aren't released back to the block cache at the end of BlockBasedTableReader::Open(). Instead the table reader takes ownership of them, hence pinning them, ie. the LRU cache will never push them out. Meanwhile in the table reader, further accesses will not hit the block cache, thus avoiding lock contention.
When the table reader is destroyed, it releases the pinned blocks (if there were any). This has to happen before the cache is destroyed, so I had to introduce a TableReader::Close(), to guarantee the order of destruction.
Test Plan:
Added two unit tests for this. Existing unit tests run fine (default is pin_l0_filter_and_index_blocks_in_cache=false).
DISABLE_JEMALLOC=1 OPT=-g make all valgrind_check -j32
Mac: OK.
Linux: with D55287 patched in it's OK.
Reviewers: sdong
Reviewed By: sdong
Subscribers: andrewkr, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D54801
Summary:
In block based table reader, wow we put index reader to block cache, which can be retrieved after DB restart. However, index reader may reference internal comparator, which can be destroyed after DB restarts, causing problems.
Fix it by making cache key identical per table reader.
Test Plan: Add a new test which failed with out the commit but now pass.
Reviewers: IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: maro, yhchiang, kradhakrishnan, leveldb, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D55287
Summary:
Cache to have an option to fail Cache::Insert() when full. Update call sites to check status and handle error.
I totally have no idea what's correct behavior of all the call sites when they encounter error. Please let me know if you see something wrong or more unit test is needed.
Test Plan: make check -j32, see tests pass.
Reviewers: anthony, yhchiang, andrewkr, IslamAbdelRahman, kradhakrishnan, sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D54705
Summary: We want to provide a way to detect whether an iterator is stale and needs to be recreated. Add a iterator property to return version number.
Test Plan: Add two unit tests for it.
Reviewers: IslamAbdelRahman, yhchiang, anthony, kradhakrishnan, andrewkr
Reviewed By: andrewkr
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D54921
Summary:
Add Iterator::GetProperty(), a way for users to communicate with iterator, and turn Iterator::IsKeyPinned() with it.
As a follow-up, I'll ask a property as the version number attached to the iterator
Test Plan: Rerun existing tests and add a negative test case.
Reviewers: yhchiang, andrewkr, kradhakrishnan, anthony, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D54783
Summary:
This reverts commit 73c31377bb, which mistakenly
reverts 73c31377bb that fixes a bug when both
whole_key_filtering and prefix_extractor are set
Test Plan: revert the patch
Reviewers: anthony, IslamAbdelRahman, rven, kradhakrishnan, sdong
Reviewed By: sdong
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D52707
Summary:
Add kSstFileTier to ReadTier, which allows Get and MultiGet to
read only directly from SST files and skip mem-tables.
kSstFileTier = 0x2 // data in SST files.
// Note that this ReadTier currently only supports
// Get and MultiGet and does not support iterators.
Test Plan: add new test in db_test.
Reviewers: anthony, IslamAbdelRahman, rven, kradhakrishnan, sdong
Reviewed By: sdong
Subscribers: igor, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D53511
Summary: Add a new option to BlockBasedTableOptions that will allow us to change the restart interval for the index block
Test Plan: unit tests
Reviewers: yhchiang, anthony, andrewkr, sdong
Reviewed By: sdong
Subscribers: march, dhruba
Differential Revision: https://reviews.facebook.net/D53721
Summary: Similar to D53385 we need to check InDomain before checking the filter block.
Test Plan: unit tests
Reviewers: yhchiang, rven, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D53421
Summary:
Right now when we are creating a BlockBasedTable with fill filter block
we add to the filter all the prefixes that are InDomain() based on the prefix_extractor
the problem is that when we read a key from the file, we check the filter block for the prefix whether or not it's InDomain()
Test Plan: unit tests
Reviewers: yhchiang, rven, anthony, kradhakrishnan, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D53385
Summary:
SstFileWriter may create an sst file with no entries
Right now this will fail when being ingested using DB::AddFile() saying that the keys are corrupted
Test Plan: make check
Reviewers: yhchiang, rven, anthony, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D52815
Summary:
After introducing Iterator::PinData(), we have extra overhead of deleting the pinned iterators that we track in a std::set
This patch avoid inserting to the std::set if we have only one iterator (normal use case when no iterators are pinned)
Before this change
```
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="newiterator" --db="/tmp/rocksdbtest-8616/dbbench" --use_existing_db --disable_auto_compactions
newiterator : 1.006 micros/op 994013 ops/sec;
newiterator : 0.994 micros/op 1006295 ops/sec;
newiterator : 0.990 micros/op 1010422 ops/sec;
```
After change
```
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="newiterator" --db="/tmp/rocksdbtest-8616/dbbench" --use_existing_db --disable_auto_compactions
newiterator : 0.754 micros/op 1326588 ops/sec;
newiterator : 0.759 micros/op 1317394 ops/sec;
newiterator : 0.691 micros/op 1446704 ops/sec;
```
Test Plan: make check -j64
Reviewers: yhchiang, rven, anthony, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D52761
Summary: <array> is not included in table/plain_table_key_coding.h. It may be the cause of one CLANG build failure.
Test Plan: Build it
Reviewers: yhchiang, rven, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D52725
Summary: In plain table reader's non-mmap mode, we only keep the most recent read buffer. However, for binary search, it is likely we come back to a location to read. To avoid one pread in such a case, we keep two read buffers. It should cover most of the cases.
Test Plan:
1. run tests
2. check the optimization works through strace when running
./table_reader_bench -mmap_read=false --num_keys2=1 -num_keys1=5000 -table_factory=plain_table --iterator --through_db
Reviewers: anthony, rven, kradhakrishnan, igor, yhchiang, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D51171
Summary:
table_test is failing because we are creating a temp InternalComparator
14:27:28 [ RUN ] BlockBasedTableTest.NoopTransformSeek
14:27:28 pure virtual method called
14:27:28 terminate called without an active exception
14:27:28 /bin/sh: line 7: 2346261 Aborted (core dumped) ./$t
Test Plan: make table_test -j64 && ./table_test --gtest_filter="BlockBasedTableTest.NoopTransformSeek"
Reviewers: igor, sdong, anthony, rven
Reviewed By: rven
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D52671
Jay Edgar