| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
This means keeping the old mmap around when we expand the database.
We could revert to read/write, except for platforms with incoherent
mmap (ie. OpenBSD), where we need to use mmap for all accesses.
Thus we keep a linked list of old maps, and unmap them when the last access
finally goes away.
This is required if we want ntdb_parse_record() callbacks to be able
to expand the database.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
|
|
We access the key on lookup, then access the data in the caller. It
makes more sense to access both at once. We also put in a likely()
for the case where the hash is not chained.
Before:
Adding 1000 records: 3644-3724(3675) ns (129656 bytes)
Finding 1000 records: 1596-1696(1622) ns (129656 bytes)
Missing 1000 records: 1409-1525(1452) ns (129656 bytes)
Traversing 1000 records: 1636-1747(1668) ns (129656 bytes)
Deleting 1000 records: 3138-3223(3175) ns (129656 bytes)
Re-adding 1000 records: 3278-3414(3329) ns (129656 bytes)
Appending 1000 records: 5396-5529(5426) ns (253312 bytes)
Churning 1000 records: 9451-10095(9584) ns (253312 bytes)
smbtorture results (--entries=1000)
ntdb speed 183881-191112(188223) ops/sec
After:
Adding 1000 records: 3590-3701(3640) ns (129656 bytes)
Finding 1000 records: 1539-1605(1566) ns (129656 bytes)
Missing 1000 records: 1398-1440(1413) ns (129656 bytes)
Traversing 1000 records: 1629-2015(1710) ns (129656 bytes)
Deleting 1000 records: 3118-3236(3163) ns (129656 bytes)
Re-adding 1000 records: 3235-3355(3275) ns (129656 bytes)
Appending 1000 records: 5335-5444(5385) ns (253312 bytes)
Churning 1000 records: 9350-9955(9494) ns (253312 bytes)
smbtorture results (--entries=1000)
ntdb speed 180559-199981(195106) ops/sec
|
|
TDB2 started with a top-level hash of 1024 entries, divided into 128
groups of 8 buckets. When a bucket filled, the 8 bucket group
expanded into pointers into 8 new 64-entry hash tables. When these
filled, they expanded in turn, etc.
It's a nice idea to automatically expand the hash tables, but it
doesn't pay off. Remove it for NTDB.
1) It only beats TDB performance when the database is huge and the
TDB hashsize is small. We are about 20% slower on medium-size
databases (1000 to 10000 records), worse on really small ones.
2) Since we're 64 bits, our hash tables are already twice as expensive
as TDB.
3) Since our hash function is good, it means that all groups tend to
fill at the same time, meaning the hash enlarges by a factor of 128
all at once, leading to a very large database at that point.
4) Our efficiency would improve if we enlarged the top level, but
that makes our minimum db size even worse: it's already over 8k,
and jumps to 1M after about 1000 entries!
5) Making the sub group size larger gives a shallower tree, which
performs better, but makes the "hash explosion" problem worse.
6) The code is complicated, having to handle delete and reshuffling
groups of hash buckets, and expansion of buckets.
7) We have to handle the case where all the records somehow end up with
the same hash value, which requires special code to chain records for
that case.
On the other hand, it would be nice if we didn't degrade as badly as
TDB does when the hash chains get long.
This patch removes the hash-growing code, but instead of chaining like
TDB does when a bucket fills, we point the bucket to an array of
record pointers. Since each on-disk NTDB pointer contains some hash
bits from the record (we steal the upper 8 bits of the offset), 99.5%
of the time we don't need to load the record to determine if it
matches. This makes an array of offsets much more cache-friendly than
a linked list.
Here are the times (in ns) for tdb_store of N records, tdb_store of N
records the second time, and a fetch of all N records. I've also
included the final database size and the smbtorture local.[n]tdb_speed
results.
Benchmark details:
1) Compiled with -O2.
2) assert() was disabled in TDB2 and NTDB.
3) The "optimize fetch" patch was applied to NTDB.
10 runs, using tmpfs (otherwise massive swapping as db hits ~30M,
despite plenty of RAM).
Insert Re-ins Fetch Size dbspeed
(nsec) (nsec) (nsec) (Kb) (ops/sec)
TDB (10000 hashsize):
100 records: 3882 3320 1609 53 203204
1000 records: 3651 3281 1571 115 218021
10000 records: 3404 3326 1595 880 202874
100000 records: 4317 3825 2097 8262 126811
1000000 records: 11568 11578 9320 77005 25046
TDB2 (1024 hashsize, expandable):
100 records: 3867 3329 1699 17 187100
1000 records: 4040 3249 1639 154 186255
10000 records: 4143 3300 1695 1226 185110
100000 records: 4481 3425 1800 17848 163483
1000000 records: 4055 3534 1878 106386 160774
NTDB (8192 hashsize)
100 records: 4259 3376 1692 82 190852
1000 records: 3640 3275 1566 130 195106
10000 records: 4337 3438 1614 773 188362
100000 records: 4750 5165 1746 9001 169197
1000000 records: 4897 5180 2341 83838 121901
Analysis:
1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%.
2) TDB starts to lose when hash chains get 10 long (fetch 10% slower
than TDB2/NTDB).
3) TDB does horribly when hash chains get 100 long (fetch 4x slower
than NTDB, 5x slower than TDB2, insert about 2-3x slower).
4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger
than TDB1
|
|
We also split off the NTDB_CONVERT case (where the ntdb is of a
different endian) into its own io function.
NTDB speed:
Adding 10000 records: 3894-9951(8553) ns (815528 bytes)
Finding 10000 records: 1644-4294(3580) ns (815528 bytes)
Missing 10000 records: 1497-4018(3303) ns (815528 bytes)
Traversing 10000 records: 1585-4225(3505) ns (815528 bytes)
Deleting 10000 records: 3088-8154(6927) ns (815528 bytes)
Re-adding 10000 records: 3192-8308(7089) ns (815528 bytes)
Appending 10000 records: 5187-13307(11365) ns (1274312 bytes)
Churning 10000 records: 6772-17567(15078) ns (1274312 bytes)
NTDB speed in transaction:
Adding 10000 records: 1602-2404(2214) ns (815528 bytes)
Finding 10000 records: 456-871(778) ns (815528 bytes)
Missing 10000 records: 393-522(503) ns (815528 bytes)
Traversing 10000 records: 729-1015(945) ns (815528 bytes)
Deleting 10000 records: 1065-1476(1374) ns (815528 bytes)
Re-adding 10000 records: 1397-1930(1819) ns (815528 bytes)
Appending 10000 records: 2927-3351(3184) ns (1274312 bytes)
Churning 10000 records: 3921-4697(4378) ns (1274312 bytes)
smbtorture results:
ntdb speed 86581-191518(175666) ops/sec
Applying patch..increase-top-level.patch
|
|
The simple "is it in range" check can be inline; complex cases can be
handed through to the normal or transaction handler.
NTDB speed:
Adding 10000 records: 4111-9983(9149) ns (815528 bytes)
Finding 10000 records: 1667-4464(3810) ns (815528 bytes)
Missing 10000 records: 1511-3992(3546) ns (815528 bytes)
Traversing 10000 records: 1698-4254(3724) ns (815528 bytes)
Deleting 10000 records: 3608-7998(7358) ns (815528 bytes)
Re-adding 10000 records: 3259-8504(7805) ns (815528 bytes)
Appending 10000 records: 5393-13579(12356) ns (1274312 bytes)
Churning 10000 records: 6966-17813(16136) ns (1274312 bytes)
NTDB speed in transaction:
Adding 10000 records: 916-2230(2004) ns (815528 bytes)
Finding 10000 records: 330-866(770) ns (815528 bytes)
Missing 10000 records: 196-520(471) ns (815528 bytes)
Traversing 10000 records: 356-879(800) ns (815528 bytes)
Deleting 10000 records: 505-1267(1108) ns (815528 bytes)
Re-adding 10000 records: 658-1681(1477) ns (815528 bytes)
Appending 10000 records: 1088-2827(2498) ns (1274312 bytes)
Churning 10000 records: 1636-4267(3785) ns (1274312 bytes)
smbtorture results:
ntdb speed 85588-189430(157110) ops/sec
|
|
This is designed to allow us to make ntdb_context (and NTDB_DATA returned
from ntdb_fetch) a talloc pointer. But it can also be used for any other
alternate allocator.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
|
|
The performance numbers for transaction pagesize are indeterminate:
larger pagesizes means a smaller transaction array, and a better
chance of having a contiguous record (more efficient for
ntdb_parse_record and some internal operations inside a transaction).
On the other hand, large pagesize means more I/O even if we change a
few bytes.
But it also controls the multiple by which we will enlarge the file,
and hence the minimum db size. It's 4k for tdb1, but 16k seems
reasonable in these modern times.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
|
|
ntdb uses tdb's transaction code, and it has an undocumented but implicit
assumption: that the transaction recovery area is always aligned to the
transaction pagesize. This means that no block will overlap the recovery
area.
This is maintained by rounding the size of the database up, so do the same
for ntdb.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
|
|
It was a hack to make compatibility easier. Since we're not doing that,
it can go away: all callers must use the return value now.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
|
|
This renames everything from tdb2 to ntdb: importantly, we no longer
use the tdb_ namespace, so you can link against both ntdb and tdb if
you want to.
This also enables building of standalone ntdb by the autobuild script.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
|
