feat(blog): Add Zihan and Ethan's final project blog #544

ethanuppal · 2025-05-14T03:50:38Z

Closes #509, closes #512

Signed-off-by: Ethan Uppal <[email protected]>

sampsyo

Hi there—looks like this is still in progress, so I won't read it yet. Please let me know when it's time to read the report.

sampsyo · 2025-05-15T01:45:35Z

content/blog/2025-05-13-zihan-ethan-project.md

+  Ethan Uppal Cornell CS '27
+  Zihan Li Cornell CS '25


Please use complete sentences.

See above; please fill in your bios.

sampsyo · 2025-05-15T01:46:35Z

content/blog/2025-05-13-zihan-ethan-project.md

+> [!NOTE]
+> Some of these questions are redundant in the context of both sections and thus their answers will be too.


Do not structure your blog post as a question-and-answer list. Remember that the audience is external: you need to write something that will be intelligible to someone who wants to learn about your project "from scratch."

sampsyo

Nice work on the overall design & implementation here! It's cool that you were able to observe nontrivial speedups for one analysis. I think it would wonderful to add some additional reflection about what you think the results mean, and what this tells us about the potential for parallelizing dataflow analyses in general.

sampsyo · 2025-05-15T19:17:25Z

content/blog/2025-05-13-zihan-ethan-project.md

+  Ethan Uppal Cornell CS '27
+  Zihan Li Cornell CS '25


See above; please fill in your bios.

sampsyo · 2025-05-15T19:18:02Z

content/blog/2025-05-13-zihan-ethan-project.md

+	if out[b] changed:
+		Worklist += successors of b
+```
+In this [project](https://github.com/zihan0822/para-dflow), we built a parallel dataflow solver in Rust with bitset optimizations for our flattened Bril IR. We parallelized the KILL and GEN set computation and the condensed cfg traversal process. We focused on one forward pass analysis: reaching definition and one backward pass analysis: liveness analysis in particular. 


sampsyo · 2025-05-15T19:18:32Z

content/blog/2025-05-13-zihan-ethan-project.md

+	if out[b] changed:
+		Worklist += successors of b
+```
+In this [project](https://github.com/zihan0822/para-dflow), we built a parallel dataflow solver in Rust with bitset optimizations for our flattened Bril IR. We parallelized the KILL and GEN set computation and the condensed cfg traversal process. We focused on one forward pass analysis: reaching definition and one backward pass analysis: liveness analysis in particular. 


We focused on one forward pass analysis: reaching definition and one backward pass analysis: liveness analysis in particular.

To make this legible, try commas or parentheses:

We focused on one forward pass analysis (reaching definitions) and one backward pass analysis (liveness analysis) in particular.

sampsyo · 2025-05-15T19:19:37Z

content/blog/2025-05-13-zihan-ethan-project.md

+
+## Preparations
+#### Flattened Bril Representation
+We implemented a flattened representation for Bril to get rid of fragmented heap references in previous Bril representations implemented in [bril-rs](https://github.com/sampsyo/bril/tree/main/bril-rs). Here are some of our flattened equivalents. 


No need to pick on bril-rs. You can just say that you created a flattened representation that avoided the heap fragmentation that can come with a standard, pointer-based program representation.

sampsyo · 2025-05-15T19:20:41Z

content/blog/2025-05-13-zihan-ethan-project.md

+}
+```
+
+With this flattened representation, we hope to isolate the performance increase to just the dataflow analyses. It also simplifies things by tying all references’ lifetime to the program. We also provide a handy shim that transforms bril’s official repr to our flattened repr. 


bril -> Bril
repr -> representation

sampsyo · 2025-05-15T20:11:03Z

content/blog/2025-05-13-zihan-ethan-project.md

+
+Both `GEN[b]` and `KILL[b]` only depend on block local info. 
+
+We parallelize KILL and GEN computation with [rayon's par_iter](https://docs.rs/rayon/latest/rayon/). 


Again, parallelize over what? Are we parallelizing over basic blocks (and then scanning the instructions within each block sequentially), or are we parallelizing over the instructions within a block?

sampsyo · 2025-05-15T20:13:18Z

content/blog/2025-05-13-zihan-ethan-project.md

+
+
+
+##### 2. Condensed CFG traversal in parallel:


Finding the SCCs and parallelizing across them is a good idea! Nice!

Do you do this for the sequential version too, or just the parallel version? It would be interesting to try both, i.e., to compare three treatments: "standard" sequential, sequential with SCCs, and parallel with SCCs.

sampsyo · 2025-05-15T20:13:53Z

content/blog/2025-05-13-zihan-ethan-project.md

+
+
+## Evaluations
+To test the correctness, we compare the results of sequential and parallel solver on core benchmarks and fuzzed programs to make sure they agree. 


What were the results?

sampsyo · 2025-05-15T20:14:20Z

content/blog/2025-05-13-zihan-ethan-project.md

+bril-fuzzer –-num-block 1024 –-block-size-mean 128 –-max-nesting 3
+```
+
+The sequential baseline is somewhat parallelized with SIMD accelerated bitset implementation. 


Can you say something more about your experimental setup? Some data that would be useful include hardware details, OS versions, Rust versions, etc., and especially the number of cores in your machine.

sampsyo · 2025-05-15T20:16:56Z

content/blog/2025-05-13-zihan-ethan-project.md

+**Liveness Analysis**: 1.85x faster 
+| Method     | Fastest (ms) | Slowest (ms) | Mean (ms) |
+|------------|--------------|---------------|-----------|
+| Parallel   | 231.6        | 233.9         | 232.7     |
+| Sequential | 427.0        | 434.2         | 430.6     |
+
+
+**Reaching Def**: 8% slow down
+| Method     | Fastest (s) | Slowest (s) | Mean (s) |
+|------------|--------------|---------------|-----------|
+| Parallel   | 17.4        | 24.11         | 20.76     |
+| Sequential | 18.76        | 19.41         | 19.08     |


Can you say something about why you think the results turned out this way? Is it due to the profile imbalance you mention below, or something else?

How about any theories for where this might go in the future? Do you think this is a promising approach that could work for other analyses, or did you learn that this is a bad idea and we should stop here? It would be great to do a little reflection about what you think these results tell you, qualitatively speaking.

Signed-off-by: Ethan Uppal <[email protected]>

…ppal/cs6120-fork-ignore into zihan-ethan-final-project

zihan0822 · 2025-05-17T20:21:28Z

We updated the evaluation setup for reaching definitions, now we can have a consistent 1.2x plus speed up with the parallel solver.  After profiling, we realized that the main bottleneck for reaching definitions was computing DEFS, which we somehow can not find an effective parallel solution for (discussed a bit more in the blog).  The performance gain by parallelizing GEN and dataflow was overshadowed by this bottleneck. 

The main modifications we made here were:

Instead of tracking the instruction offset for each definition, we only tracked the block id associated with it. This reduced the memory footprint by a factor of num_total_instructions / num_blocks.
We allocated a bitset arena to serve the frequent bitset allocation requests.

Those modifications were applied to both the sequential and parallel solver.

sampsyo · 2025-05-18T13:59:04Z

Wonderful! This is looking great. Seriously impressive work here.

feat(blog): Add Zihan and Ethan's WIP final project blog

59ee7b8

Signed-off-by: Ethan Uppal <[email protected]>

ethanuppal changed the title ~~feat(blog): Add Zihan and Ethan's WIP final project blog~~ feat(blog): Add Zihan and Ethan's final project blog May 14, 2025

sampsyo reviewed May 15, 2025

View reviewed changes

zihan0822 added 2 commits May 15, 2025 10:40

zihan-ethan parallel dataflow project blog draft

27fb565

fix minor formatting issue

7121d83

zihan0822 force-pushed the zihan-ethan-final-project branch from 24b7f93 to 7121d83 Compare May 15, 2025 14:51

sampsyo requested changes May 15, 2025

View reviewed changes

zihan0822 added 5 commits May 15, 2025 17:08

fix trivial wording issue

59431e5

clarify parallel KILL GEN computation

72c1c25

add more comments on evaluation sections and future work

49145eb

add brief descriptions of flattened Bril representation

c84c2b1

add full bio

2b8d9d7

ethanuppal force-pushed the zihan-ethan-final-project branch 6 times, most recently from 21721f8 to 9218138 Compare May 16, 2025 04:01

refactor(blog): Improve and rewrite for clarity, formatting, and intent

dd3476d

Signed-off-by: Ethan Uppal <[email protected]>

ethanuppal force-pushed the zihan-ethan-final-project branch 2 times, most recently from c26566d to 61bbcc5 Compare May 16, 2025 04:13

fix: Command was listed incorrectly

4db6a5f

Signed-off-by: Ethan Uppal <[email protected]>

ethanuppal force-pushed the zihan-ethan-final-project branch from 61bbcc5 to 4db6a5f Compare May 16, 2025 04:13

zihan0822 and others added 3 commits May 16, 2025 01:48

fix a couple of typos

7a1a179

feat: Explain code listing for flattened repr

3ff821d

Signed-off-by: Ethan Uppal <[email protected]>

Merge branch 'zihan-ethan-final-project' of https://github.com/ethanu…

6baf2cc

…ppal/cs6120-fork-ignore into zihan-ethan-final-project

sampsyo added the 2025sp label May 16, 2025

zihan0822 added 2 commits May 16, 2025 20:05

update experiment setup and evaluation results

452ad71

add comments on precomputation free alternative

1ffa48c

sampsyo merged commit 69ca357 into sampsyo:2025sp May 18, 2025
2 checks passed

		> [!NOTE]
		> Some of these questions are redundant in the context of both sections and thus their answers will be too.


		Both `GEN[b]` and `KILL[b]` only depend on block local info.

		We parallelize KILL and GEN computation with [rayon's par_iter](https://docs.rs/rayon/latest/rayon/).



		## Evaluations
		To test the correctness, we compare the results of sequential and parallel solver on core benchmarks and fuzzed programs to make sure they agree.

feat(blog): Add Zihan and Ethan's final project blog #544

feat(blog): Add Zihan and Ethan's final project blog #544

Uh oh!

Conversation

ethanuppal commented May 14, 2025

Uh oh!

sampsyo left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

sampsyo left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

zihan0822 commented May 17, 2025

Uh oh!

sampsyo commented May 18, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants