Write-optimization in external memory data structures (Highload++ 2014)

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Write-optimization in external memory data structures
Leif Walsh
Tokutek, Inc.
leif@tokutek.com
@leifwalsh
November 1, 2014
Leif Walsh (Tokutek) Fractal Trees November 1, 2014 1 / 31

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Background

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Data structures:

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Data structures:
Provide retrieval of data.
Lookup(Key)
Pred(Key)
Succ(Key)

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Data structures:
Provide retrieval of data.
Lookup(Key)
Pred(Key)
Succ(Key)
Dynamic data structures let you change
the data.
Insert(Key; Value)
Delete(Key)

.
.
.
.
.
.
.
.
.
[Aggarwal & Vitter ’88]
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
DAM model
Problem size N.
Memory size M.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
DAM model
Problem size N.
Memory size M.
Transfer data to/from memory in blocks
of size B.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
DAM model
Problem size N.
Memory size M.
Transfer data to/from memory in blocks
of size B.
Efficiency of operations is measured as the
number of block transfers, a.k.a. IOPS.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
A B-tree (Б-tree?) is an external memory data structure:

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Balanced search tree.
Fanout of B
(block size / key size).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Fanout of B
Internal nodes < M.
Leaf nodes > M.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Fanout of B
Internal nodes < M.
Leaf nodes > M.
Search: O(logB N) I/Os
Insert: O(logB N) I/Os

.
.
.
.
.
.
.
[Brodal & Fagerberg ’03]
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
OLAP

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Write-optimization technique #1: OLAP
OLAP: Online Analytical Processing

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Key idea: Analyze data collected in the past.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Key idea: Analyze data collected in the past.
B-tree inserts are slow, but…logging and sorting are fast.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Merge sort:

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Merge sort in external memory:

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Merge sort cost in DAM model is:
Cost to scan through all the data once.
Multiplied by the # of levels in the merge tree.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
N/B

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
N/B
logM/B N/B

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
N/B
logM/B N/B
O
(
N
B
logM/B
N
B
)

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Insert N elements into a B-tree:
O
(
N logB
N
M
)
Merge sort:
O
(
N
B
logM/B
N
B
)

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
O
(
N logB
N
M
)
Merge sort:
O
(
N
B
logM/B
N
B
)
2N
B
Typically, M/B is large, so only two passes are needed to sort.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
O
(
N logB
N
M
)
N
Merge sort:
O
(
N
B
logM/B
N
B
)
2N
B
Intuition: Each insert into a B-tree costs 1 seek, while sorting is close to disk bandwidth.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Insert N elements into a B-tree: (assuming 100-1000 byte elements)
O
(
N logB
N
M
)
N 10 100kB/s = 100 elements/s
Merge sort:
O
(
N
B
logM/B
N
B
)
2N
B

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Insert N elements into a B-tree: (assuming 100-1000 byte elements)
O
(
N logB
N
M
)
N 10 100kB/s = 100 elements/s
Merge sort:
O
(
N
B
logM/B
N
B
)
2N
B
50MB/s = 50k 500k elements/s

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
So, how does OLAP work?

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Log new data unindexed until you accumulate a lot of it (10% of the data set).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Sort the new data.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Sort the new data.
Use a merge pass through existing indexes to incorporate new data.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Sort the new data.
Use indexes to do analytics.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Sort the new data.
Use indexes to do analytics.
Moral: OLAP techniques can handle high insertion volume, but query results are delayed.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
LSM-trees

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Write-optimization technique #2: LSM-trees
The insight for LSM-trees starts by asking: how can we reduce the queryability delay in OLAP?

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
The buffer is small, let’s index it!

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Inserts go into the “buffer B-tree”.
When the buffer gets full, we merge it with the “main B-tree”.
Queries have to touch both trees and merge results, but results are available immediately.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Inserts go into the “buffer B-tree”.
When the buffer gets full, we merge it with the “main B-tree”.
Queries have to touch both trees and merge results, but results are available immediately.
(This specific technique (which is not yet an LSM-tree) is used in InnoDB and is called the “change buffer”.)

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Why is this fast?

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Why is this fast?
The buffer is in-memory, so inserts are fast.
When we merge, we put many new elements in each leaf in the main B-tree (this amortizes
the I/O cost to read the leaf ).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Why is this fast?
Eventually, we reach a problem:

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Why is this fast?
Eventually, we reach a problem:
If the buffer gets too big, inserts get slow.
If the buffer stays too small, the merge gets inefficient because each leaf node receives
only a few elements (back to O(N logB N)).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
How can we fix this?

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
How can we fix this? More buffering!

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Each level is twice as large as the previous level, for some value of 2 (usually 10).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Each level is twice as large as the previous level, for some value of 2 (usually 10). We’ll use 2.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
How do queries work?

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Search cost is:
logB B + : : : + logB
N
8
+ logB
N
4
+ logB
N
2
+ logB N

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Search cost is:
N
8
+ logB
N
4
+ logB
N
2
+ logB N
=
1
log B (1 + : : : + lg(N) 3 + lg(N) 2 + lg(N) 1 + lg(N))

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Search cost is:
N
8
+ logB
N
4
+ logB
N
2
+ logB N
=
1
log B (1 + : : : + lg(N) 3 + lg(N) 2 + lg(N) 1 + lg(N)) = O(log N logB N)

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
How much do inserts cost?

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Cost to flush a tree Tj of size X is O(X/B).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Cost per element to flush Tj is O(1/B).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Each element moves log N times.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Each element moves log N times.
Total amortized insert cost per element is O
(
log N
B
)
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Fractal Trees

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Write-optimization technique #3: Fractal Trees
The pain in LSM-trees is doing a full O(logB N) search in each level.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
We use fractional cascading to reduce the search per level to O(1).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
We use fractional cascading to reduce the search per level to O(1).
The idea is that once we’ve searched Ti, we know where the key would be in Ti, and we can use
that information to guide our search of Ti+1.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Add forwarding pointers from leaves in Ti to leaves in Ti+1 (but remove the redundant ones that
point to the same leaf ):

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Add ghost pointers to leaves not pointed to in Ti+1 in leaves in Ti:

[Bender, Farach-Colton, Fineman, Fogel, Kuszmaul, Nelson ’07]
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Now, after searching Ti for a missing element c, we look left and right for forwarding or ghost
pointers, and follow them down to look at O(1) leaves in Ti+1.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
This way, search is only O(logR N) (in our example, R = 2).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
This way, search is only O(logR N) (in our example, R = 2).
The internal node structure in each level is now redundant, so we can represent each level as an
array. This is called a Cache-Oblivious Lookahead Array.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Though the amortized analysis says our inserts are fast, when we flush a very large level to the
next one, we might see a big stall. Concurrent merge algorithms exist, but we can do better.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
We break each level’s array into chunks that can be flushed independently. Each chunk flushes
to a localized region of a few chunks in the next level down, found using its forwarding pointers.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Now we have a tree again!

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Now we have a tree again!
As it turns out, this structure makes it easier to manage an LRU-style cache of blocks and is more
flexible in the face of “hotspot” workloads.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Results
Modified B-tree-like dynamic (inserts, updates, deletes) data structure that supports point
and range queries.
Inserts (
are a factor B/ log B (typically 10-100x in practice) faster than a B-tree:
O
log N
B
)
O
(
log N
log B
)
.
Searches are a factor log B/ log R slower than a B-tree: O
(
log N
log R
)
O
(
log N
log B
)
.
To amortize flush costs over many elements, we want each block we write to be large
(4MB), much larger than typical B-tree blocks (16KB). These compress well.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Applications
TokuDB for MySQL, TokuMX for MongoDB:
Faster indexed insertions.
Hot schema changes.
Compression.
Faster replication on secondaries (TokuMX).
Lower impact migrations (TokuMX).
Fast (no read before write) updates (in TokuDB, coming soon in TokuMX).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Applications
TokuDB for MySQL, TokuMX for MongoDB:
Faster indexed insertions.
Hot schema changes.
Compression.
Faster replication on secondaries (TokuMX).
Lower impact migrations (TokuMX).
Fast (no read before write) updates (in TokuDB, coming soon in TokuMX).
ACID transactions.
Concurrency (TokuMX).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Benchmarks

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Questions?
Leif Walsh
leif@tokutek.com
@leifwalsh
Downloads: www.tokutek.com/downloads
Docs: docs.tokutek.com
Slides: slidesha.re/1tqwORg

Write-optimization in external memory data structures (Highload++ 2014)

Recommended

Recommended

More Related Content

Viewers also liked

Viewers also liked (15)

Recently uploaded

Recently uploaded (20)

Write-optimization in external memory data structures (Highload++ 2014)