Block Storage Performance
Benchmarking Summary
Nirvana Accelerated Block Storage (ABS) vs AWS EBS (gp3 & io2)
ClickBenchFIOPricingAssessment
Nirvana ABS delivers a clear price-performance edge, outpacing AWS gp3 on cold-read and matching io2-class behavior at far lower cost, making it a strong default for latency-sensitive, real-time workloads.
Performance Highlights
10-14x faster
on cold-read analytical queries
77% faster
on IO-heavy analytical queries
5× higher
20,000 sustained IOPS per VM (80/20 read/write)
408-430 MB
per second per VM
Methodology
Measure end-to-end analytical query performance under forced cold-read conditions.
- Benchmarks compared Nirvana Labs Accelerated Block Storage (ABS) against AWS EBS gp3 and AWS EBS io2.
- Executed 43 standardized OLAP queries (Q0–Q42) on the hits_100m dataset to measure real analytical query performance.
- OS page cache was dropped before each query, ensuring results reflect true storage performance rather than memory caching.
- Compute environments were normalized across tests with identical cores and architecture. (64 vCPU, 512 GB RAM, 25 Gbps networking, and 50 GB volumes)
- Each platform was benchmarked twice, with only the second cold-run results analyzed to ensure consistency and reproducibility.
Measure raw block storage performance independent of any application layer, isolating the storage path from database or caching effects.
- Tests simulated common infrastructure workloads to evaluate IOPS ceilings, latency across block sizes, per-VM bandwidth limits, and mixed read/write scaling behavior.
- The primary benchmark ran across a 16-VM cluster using LUNs mounted via iSCSI through the Portworx CSI driver, exercising the full production data path: VM → CSI → iSCSI → ABS backend.
- Direct I/O was enabled to bypass the OS page cache, ensuring results reflect true storage performance.
- Additional validation included post-tuning reruns for stability and single-VM isolation tests to measure raw per-node capability.
Compare monthly cost among Nirvana ABS, AWS EBS gp3, and AWS EBS io2.
- Pricing was evaluated using a normalized workload of a 1 TB block storage volume sustaining 10,000 IOPS, representative of common blockchain and database deployments.
- Monthly costs were calculated by combining storage cost, included baseline performance, and additional IOPS charges where applicable.
- The analysis focuses on effective monthly cost for sustained workloads, excluding additional compute requirements such as Nitro-based instances.
ClickBench: Cold-Read Analytics
43 OLAP queries · OS cache dropped · Real-world SQL workloads
10.5x
Median speedup vs gp3
14.1x
Speedup on IO-heavy queries
98%
Within 2x of io2 performance
77%
Within 1.5× of io2 performance
20–40×
Faster on large scan queries (Q20–Q22)
Sub-second
Cold-read scans vs gp3 expands to tens of seconds
Cold-Read Analytical Performance
ABS outperforms gp3 on cold-read analytical workloads.
Performance closely approaches io2 across most queries.
ABS maintains stable latency on large scans, avoiding gp3 performance cliffs.
CPU-bound queries converge across platforms, confirming storage drives the observed differences.
FIO: Synthetic Storage Behavior
16-VM cluster · Direct I/O · Full production path tested
Small Random Writes (≤2K)
Packet-rate bound behavior · network block storage characteristic
4K Random Writes
>20K IOPS (Free bursts allowed, short periods) · ~3.5 ms latency · Database-aligned performance
Mixed Workload (80/20)
20K sustained IOPS · Bursting enabled for short periods per VM
Sequential Throughput
Stable per-VM bandwidth ceiling · Predictable throughput
Post-Tuning Validation
Steady-state runtime improves small-IO performance
Single-VM Isolation
Backend capability verified · cluster limits reflect network path
Price-Performance Advantage
Normalized: 1 TB volume · 10,000 sustained IOPS
ABS delivers io2-class performance at ~8× lower cost.
ABS includes 20K IOPS and throughput by default, while gp3 and io2 require additional provisioning and instance requirements.
Workload Fit
Web3
RPC Infrastructure
Blockchain Foundations
Indexing/ ETL Pipelines
AI
Vector Database
LLM Inference
Agent Sandboxes
Real-time Systems
Analytics
ETL Pipelines
High Frequency Trading
Powering AI, blockchain, and
real-time systems.
