Andrei Gudkov

Software architect/researcher

Location: Moscow, Russia

LinkedIn: linkedin.com/in/andrei-gudkov/

Technical blog: gudok.xyz

I create high-performance, algorithmically-complex server software. My current focus is on scheduling algorithms for Huawei Cloud. Previous experience includes full-text search engines and a number of startups. I rely on C++/Linux as primary tech stack, followed by Python for prototyping and research.

Skills: Distributed computing, full-text search, scheduling theory, combinatorial optimization (packing problems), performance optimization (multithreading, vectorization), cloud internals, network simulation, classical ML, technical leadership.

Toolset: C++/Python/Java, Linux A-Z (I prefer Debian), ILP/SAT solvers, x86_64 SIMD, Ascend NPU, arrow/parquet, protobuf, Hadoop, HBase, JNI, fastutil, nio.

Experience

Research in the area of scheduling algorithms for Huawei Cloud (IaaS and AI training).

Created algorithms for VM consolidation, which reduce number of used hosts by up to 18.7% by live-migrating VMs. NUMA-awareness, affinity/antiaffinty rules, migration duration prediction using QEMU counters.
Created some core algorithms (scheduler, hotspot resolver, theoretical bounds) for CPU-overbooked scheduling solution with guarantees on performance SLO. Stochastic optimization, simulation of Linux CFS.
Designed constraints for scheduling specific class of LLM training jobs into proprietary FatTree-like network to ensure zero inter-job flow collisions. Reduction in training time up to 4.5%.

Publications:

"BalCon — resource balancing algorithm for VM consolidation", FGCS 2023 ↵
"Hotspot-Aware Scheduling of Virtual Machines With Overcommitment for Ultimate Utilization in Cloud Datacenters", T-ASE 2024 ↵
"Hotspot resolution in cloud computing: A Γ-robust knapsack approach for virtual machine migration", JPDC 2024 ↵

Worked remotely for bigdata and AI startups. Some interesting projects:

Lighthouse AI — smart security cameras with robust object recognition

Separate GPU nodes, data pipeline latency optimization, audio subsystem, metrics collection/monitoring.

DCEngines — low-level, distributed data store integrated with Apache Spark

String/substring indexes, parquet-based storage format, cluster API, unit and integration testing frameworks.

I also started to write job-related articles in my spare time (see my blog).

Worked on hierarchical distributed database used for risk management.

Created database client embedded into in-grid calculations (network programming, context-dependent grammars parsing, complex object model)
Implemented reference counting GC that led to x2 reduction in weekly maintenance duration
Latency/throughput/memory optimization of everything at all levels

[Go.mail.ru was the third-largest Russian web search engine with 30M search queries daily as of 2013]

Was responsible for development of all "vertical" search services. Worked in close collaboration with PM, colleagues from UI, QA, devops and other specialized groups. Major team achievements:

Created news search: 1000+ sources, clustering, archive, ≤5 min updates
Created search service over social profiles: 300M profiles, age-/geo- proximity ranking
Enriched web SERP with 15+ non-organic results: goods, images, music, news, video, weather
Personally worked on improving image search; added CTR-based ranking
Owned 100 online servers with highly heterogeneous roles
Led team of 7 developers, conducted technical interviews

Owned image search (2M queries daily).

Created indexing pipeline in Hadoop/HBase. Increased effective corpus size to 300M docs
Migrated search backend to distributed architecture with sharding, replication and thumbnail serving farm

[Nigma.Ru was a Russian metasearch engine with a variety of additional search services]

Created and owned search engine over audio files accessible on the web. This included complete dataflow and backend: robot (crawl set of 1 bln URLs), indexer, ranker, retriever, clustering and data mining algorithms, etc
Created a number of in-house components reused by adjacent projects: robot, HTML parser (30 MB/s), full-text search engine
Led team of 3 developers, conducted technical interviews

Education

Faculty of Computational Mathematics and Cybernetics
University degree in Applied Mathematics and Computer Science