Skip to content

Latest commit

 

History

History
206 lines (151 loc) · 7.25 KB

README.md

File metadata and controls

206 lines (151 loc) · 7.25 KB

HHDS: Hardware Hierarchical Dynamic Structure

This repository contains a highly optimized graph and tree data structure. The structure is optimized for handling graphs/trees structures typically found in hardware tools.

Graph

Some of the key characteristics:

-Graph mutates (add/remove edges and nodes)

-At most 32 bit ID for node (graph would be partitioned if larger is needed)

-Once an ID is created, even if node is deleted, it can not be reused. This sometimes called a "tombstone deletion"

-Delete node is frequent (code optimization)

-2 main types of traversal (any order or topological sort)

-Add edges is not so frequent after the 1st phase of graph creation

-Nodes has several "pins" and the edges are bi-directional

-In practice (topological sort), most edges are "short" no need to keep large 32 bit integer. Delta optimization to fit more edges.

-Meta information (attributes) are kept separate with the exception of type

-Most nodes have under 8 edges, but some (reset/clk) have LOTS of edges

-The graph operations do NOT need to be multithreaded with updates. The parallelism is extracted from parallel thread operations. (parallel read-only may happen).

Related works

Some related (but different) graph representations:

Pandey, Prashant, et al. "Terrace: A hierarchical graph container for skewed dynamic graphs." Proceedings of the 2021 International Conference on Management of Data. 2021.

Winter, Martin, et al. "faimGraph: high performance management of fully-dynamic graphs under tight memory constraints on the GPU." SC18: International Conference for High Performance Computing, Networking, Storage and Analysis. IEEE, 2018.

Bader, David A., et al. "Stinger: Spatio-temporal interaction networks and graphs (sting) extensible representation." Georgia Institute of Technology, Tech. Rep (2009).

Feng, Guanyu, et al. "Risgraph: A real-time streaming system for evolving graphs to support sub-millisecond per-update analysis at millions ops/s." Proceedings of the 2021 International Conference on Management of Data. 2021.

Overall data structure:

Vector with 4 types of nodes: Free, Node, Pin, Overflow

Node/Pin are 16 byte, Overflow 32 byte

Node is the master node and also the output pin

Pin is either of the other node pins (input 0, output 1....)

Node/Pin have a very small amount of edges. If more are needed, the overflow is used. If more are needed an index to emhash8::HashSet is used (overflow is deleted and moved to the set)

The overflow are just a continuous (not sorted) array. Scanning is needed if overflow is used.

Tree

The tree data structure is optimized for typical AST traversal and operations. Once a tree is created, the common operation is to delete nodes, and at most insert phi nodes (SSA). This means that the insertion is for a few entries, not large subtrees.

The traversal is a topological sort equivalent where parents of leafs are visited first, and then to process them the leafs are accessed. In a way, the following is a typical AST traversal:

Index: first Child, parent

01: 00,03 │   +── 1.1.1
02: 00,03 │   |── 1.1.2
03: 01,32 -── 1.1
04: 00,06 │           +── 1.2.1.1.1
05: 00,06 │           |── 1.2.1.1.2
06: 04,13 │       +── 1.2.1.1
07: 00,08 │           +── 1.2.1.2.1
08: 07,13 │       |── 1.2.1.2
09: 00,12 │           +── 1.2.1.3.1
10: 00,12 │           |── 1.2.1.3.2
11: 00,12 │           |── 1.2.1.3.3
12: 09,13 │       |── 1.2.1.3
13: 04,27 │   +── 1.2.1
14: 00,16 │       +── 1.2.2.1
15: 00,16 │       |── 1.2.2.2
16: 14,27 │   ├── 1.2.2
17: 00,19 │           +── 1.2.3.1.1
18: 00,19 │           |── 1.2.3.1.2
19: 17,23 │       +── 1.2.3.1
20: 00,22 │           +── 1.2.3.2.1
21: 00,22 │           |── 1.2.3.2.2
22: 20,23 │       +── 1.2.3.2
23: 19,27 │   |── 1.2.3
24: 00,26 │       +── 1.2.4.1
25: 00,26 │       |── 1.2.4.2
26: 24,27 │   |── 1.2.4
27: 13,32 ├── 1.2
28: 00,30 │   |── 1.3.1
29: 00,30 │   |── 1.3.2
30: 00,32 ├── 1.3
31: 00,32 ├── 1.4
32: 00,00 | 1

An alternative data structure.

Index: last_child, parent
// 01: 16,00 | 1
// 02: 00,01 -── 1.1
// 03: 04,01 ├── 1.2
// 04: 05,03 │   -── 1.2.1
// 05: 00,04 │       -── 1.2.1.1
// 06: 00,04 │       ├── 1.2.1.2
// 07: 10,01 ├── 1.3
// 08: 00,07 │   -── 1.3.1
// 09: 00,07 │   ├── 1.3.2
// 10: 13,07 │   ├── 1.3.3
// 11: 00,10 │       -── 1.3.1.1
// 12: 00,10 │       |── 1.3.1.2
// 13: 15,10 │       |── 1.2.1.3
// 14: 00,13 │           -── 1.2.1.3.1
// 15: 00,13 │           -── 1.2.1.3.2
// 16: 20,01 ├── 1.4
// 17: 00,16 │   -── 1.4.1
// 18: 19,16 │   ├── 1.4.2
// 19: 00,18 │   │   -── 1.4.2.1
// 20: 21,16 │   ├── 1.4.3
// 21: 00,20 │   │   -── 1.4.3.1

The previous data structure fits quite well with most AST (LiveHD) tree operations with the exception of two popular ones: get_next_sibling and add_child.

An alternative data structure. The siblings are consecutive, the descendent from the siblings are after all the siblings.

Index: first_child, last_child, parent
// 01: 02,05,00 | 1
// 02: 00,00,01 -── 1.1
// 03: 06,06,01 ├── 1.2
// 04: 12,14,01 ├── 1.3
// 05: 17,19,01 ├── 1.4
// 06: 07,09,03 │   -── 1.2.1
// 07: 00,00,06 │       -── 1.2.1.1
// 08: 00,00,06 │       ├── 1.2.1.2
// 09: 10,11,06 │       |── 1.2.1.3
// 10: 00,00,09 │           -── 1.2.1.3.1
// 11: 00,00,09 │           -── 1.2.1.3.2
// 12: 15,16,04 │   -── 1.3.1
// 13: 00,00,04 │   ├── 1.3.2
// 14: 00,00,04 │   ├── 1.3.3
// 15: 00,00,12 │       -── 1.3.1.1
// 16: 00,00,12 │       |── 1.3.1.2
// 17: 00,00,05 │   -── 1.4.1
// 18: 20,21,05 │   ├── 1.4.2
// 19: 00,00,05 │   ├── 1.4.3
// 20: 00,00,18 │   │   -── 1.4.2.1
// 21: 00,00,18 │   │   -── 1.4.3.1

The previous structure handles everything quite fast with the exception of add_child

add_child tends to add nodes close to the end. In which case, and append or just shifting a few nodes works fast.

Some solutions for random location add_child:

-Support a small hashmap with "other children" that did not fit. Only insert starting from last_child, and change last_child to 0 to indicate that last children are in a hashmap.

-Allow it BUT trigger a message to fix. This should be rare and easy to fix so that calls are not in that order

-Chunk the vector to blocks of 16K entries. This allows smaller pointers (first_child,last_child, parent) using shorts (16bits). WHen a pointer needs to point to other table (16K entry max per table), it can have a all ones value (-1) to indicate that the following entry uses the 3 fields (16 bits) as pointer to the larger table (32 bits) and 16 bits as offset.

Related

Not same, but similar idea and different representation:

Meyerovich, Leo A., Todd Mytkowicz, and Wolfram Schulte. "Data parallel programming for irregular tree computations." (2011).

Vollmer, Michael, Sarah Spall, Buddhika Chamith, Laith Sakka, Chaitanya Koparkar, Milind Kulkarni, Sam Tobin-Hochstadt, and Ryan R. Newton. "Compiling tree transforms to operate on packed representations." (2017): 26.