On GitHub, I maintain or contribute to projects such as fxamacker/cbor, fxamacker/circlehash, onflow/atree, onflow/ccf, onflow/cadence, onflow/flow-go, etc.

My first open source project is fxamacker/cbor. fxamacker/cbor is used in projects by Arm Ltd., Cisco, Dapper Labs, EdgeX�?Foundry, Fraunhofer‑AISEC, Linux�?Foundation, Microsoft, Mozilla, Tailscale, Teleport, and�?others.

fxamacker/cbor passed multiple confidential security assessments in 2022. A nonconfidential security assessment (prepared by NCC�?Group for Microsoft�?Corporation) includes a subset of fxamacker/cbor v2.4.0 without finding any vulnerabilities.

Most of the code I wrote is closed source (in many languages but mostly multithreaded C++). I'm currently enjoying open source projects and the amazing simplicity of Go.

Some of my open source work is described here.

Design & Implementation

onflow/atree: Atree provides scalable arrays and maps. It is used by Cadence in the Flow Blockchain.

Atree segments, encodes, and stores data into chunks of relatively small size. This enables blockchains to only hash and transmit modified chunks (aka payloads) instead of the entire array, map, or large element.

Among other aspects, I designed and implemented a novel hash collision handling method. It is different from published methods such as Cuckoo Hashing, Double Hashing, 2-Choice Hashing, etc.

This new hash collision handling method balances speed, security, storage size, etc. It uses a fast noncryptographic 64-bit hash and if there is a hash collision, it uses deferred and segmented 256-bit cryptographic digest (in 64-bit segments). By default, it uses CircleHash64f and BLAKE3.

Acknowledgements: Atree wouldn't exist without Dieter Shirley making priorities clear and inspiring us, Ramtin M. Seraj leading the R&D and empowering us to innovate, and Bastian Müller improving Atree while leading the integration into Cadence. Many thanks to Supun Setunga for the very complex data migration work and more!

Optimizations

When feasible, my optimizations improve speed, memory, storage, and network use without negative tradeoffs.

onflow/flow-go: Found optimizations by reading unfamiliar source code and proposed them to resolve issue #1750. Very grateful for Ramtin M. Seraj for opening a batch of issues and letting me tackle this one.

PR #1944 (Optimize MTrie Checkpoint for speed, memory, and file size):

• SPEED: 171x speedup (11.4 hours to 4 minutes) in MTrie traversing/flattening/writing phase (without adding concurrency) which led to a 47x speedup in checkpointing (11.7 hours to 15 mins).
• MEMORY: -431 GB alloc/op (-54.35%) and -7.6 billion allocs/op (-63.67%)
• STORAGE: -6.9 GB file size (without using compression yet)

After PR #1944 reduced Mtrie flattening and serialization phase to under 5 minutes (which sometimes took 17 hours on mainnet16), creating a separate MTrie state used most of the remaining duration and memory.

Additional optimizations (add concurrency, add compression, etc.) were moved to separate issue/PR and I switched my focus to related issues like #1747.

UPDATE: About six months later, file size grew from 53GB to 126GB and checkpointing frequency increased to every few hours (instead of about once daily) due to increased transactions and data size. Without PR #1944, checkpointing would be taking over 20-30 hours each time, require more operational RAM, and slowdown the system with increased gc pressure. More info: issue #2286 and PR #2792.

Evaluations and Improvements

fxamacker/circlehash: I created CircleHash64 on weekends after evaluating state-of-the-art fast hashes for work. At the time, I needed a fast hash for short input sizes typically <128 bytes but didn't like existing hashes. I didn't want to reinvent the wheel so I based it on Google Abseil C++ internal hash. CircleHash64 is well-rounded: it balances speed, digest quality, and maintainability.

CircleHash64 has good results in Strict Avalanche Criterion (SAC).

☝️ Using demerphq/smhasher updated to test all input sizes 0-128 bytes (SAC test will take hours longer to run).

CircleHash64 is fast at hashing short inputs with a 64-bit seed

• Using Go 1.17.7, darwin_amd64, i7-1068NG7 CPU
• Results from go test -bench=. -count=20 and benchstat
• Fastest XXH64 in Go+Assembly doesn't support seed

CircleHash64 doesn't have big GB/s drops in throughput as input size gets larger. Other CircleHash variants are faster for larger input sizes and a bit slower for short inputs (not yet published).

Implement IETF Internet Standards (RFC 8949 & RFC 7049)

fxamacker/cbor: I designed and implemented a secure CBOR codec after reading RFC 7049. During implementation, I helped review the draft leading to RFC 8949. The CBOR codec rejects malformed CBOR data and has an option to detect duplicate map keys. It doesn't crash when decoding bad CBOR data.

Decoding 9 or 10 bytes of malformed CBOR data shouldn't exhaust memory. For example,
[]byte{0x9B, 0x00, 0x00, 0x42, 0xFA, 0x42, 0xFA, 0x42, 0xFA, 0x42}

*Speed and memory are for latest codec version listed in the row (compiled with Go 1.17.5).

fxamacker/cbor CBOR safety settings include: MaxNestedLevels, MaxArrayElements, MaxMapPairs, and IndefLength.

Professional Background

I try to balance competing factors such as speed, security, usability, and maintainability based on each project's priorities.

Most recently, I accepted an offer I received on April 13, 2021 from Dapper Labs. I had been working for them as an independent contractor for about two weeks to help optimize Cadence storage layer and to create a streaming mode branch of fxamacker/cbor. On my first day as a contractor, I created issue 738 and the Cadence team was very welcoming and productive to work with. I subsequently opened 100+ issues and 100+ PRs at work in 2021.

My prior experience before Dapper Labs includes co-founding & bootstrapping enterprise software company, and working as an IT consultant.

• My smallest consulting client - a startup. I assisted with prototyping which helped secure their next round of funding.
• My largest consulting client - an S&P 500 company with almost 50,000 employees. I evaluated (as part of a large team) various technologies to be selected for their new global stack for deployment to over 100 countries.
• My largest software licensing+subscription+support client - a company with over 3,000 employees in the US that deployed my data security system to all their US-based offices and factories. The tamper-resistant system used 4 types of servers to distribute and enforce security policies across multiple timezones for various client software. The system was designed to repair and update itself with bugfixes without introducing downtime. I was only one of two people to ever have access to the source code: just two of us conceived, designed, implemented, tested, and maintained the system. Our system beat enterprise solutions from well-funded large competitors year after year during customer evaluations which included testing employee-attempted data theft. It was not approved for export or use outside the US.

Developing commercial software provided the advantage of choosing the most appropriate language and framework for each part of the system because the customers didn't know what programming languages, tools, or frameworks were used.