How Alluxio Boosts Tencent Cloud EMR: Cutting Bandwidth by 50% and Accelerating IO‑Intensive Workloads

1. Current Big Data Challenges

Rapid growth of data volume (PB to EB) creates data silos, rigid scaling, low resource utilization, and job congestion when massive datasets are processed concurrently. Traditional integrated compute‑storage clusters struggle to meet elastic demand and incur high OPEX.

2. Tencent Cloud Elastic MapReduce (EMR)

EMR now supports three storage back‑ends: EMR‑HDFS, EMR‑COS, and EMR‑CHDFS. EMR‑COS and EMR‑CHDFS provide out‑of‑the‑box compute‑storage separation, enabling on‑demand compute while keeping storage independent.

EMR‑HDFS : storage size tied to cluster scale.

EMR‑COS : massive, low‑cost object storage.

EMR‑CHDFS : massive, high‑performance HDFS‑compatible storage.

3. Optimizing Compute‑Storage Separation with Alluxio

By collaborating with the Alluxio community, the EMR team incorporated Alluxio 2.3.0 to address three main pain points:

Memory‑level I/O : Alluxio acts as a distributed cache, delivering memory‑speed reads for hot data and leveraging tiered storage (memory, SSD, disk).

Improved data locality : Deploying Alluxio workers alongside compute nodes allows direct memory‑level access, reducing remote fetches.

Simplified cloud/object storage access : Alluxio abstracts differing semantics of COS and CHDFS, avoiding costly metadata operations and providing unified namespace.

Additional benefits include single‑point access to heterogeneous data sources and reduced management complexity.

4. Performance Evaluation and Tuning

Benchmarks were conducted with TPC‑DS on Spark using EMR‑2.5.0 (1 Master + 25 Core nodes). The test suite measured bandwidth usage and query latency.

4.1 Bandwidth Reduction

Results show a 20‑50% reduction in peak bandwidth and a 10‑50% decrease in total bandwidth consumption.

4.2 Query Performance

Across most scenarios, especially I/O‑intensive workloads, query execution time improved by 5‑40%.

4.3 Targeted Optimizations

Data locality : Co‑locating Alluxio workers with compute nodes and tuning policies such as block.read.location.policy and writetype.default.

Metadata tuning : Leveraging Alluxio’s Catalog Service and adjusting path.caching.thread, path.cache.capacity, and inode handling to mitigate metadata bloat.

Java GC mitigation : Integrating Tencent Kona JDK to improve GC scheduling and memory release for the Alluxio Java process.

5. Conclusions

The Alluxio‑enhanced EMR solution effectively lowers bandwidth costs, accelerates I/O‑heavy jobs, and maintains elastic scalability, making it a compelling choice for enterprises adopting compute‑storage separation in the cloud.