Benchmarking Block Storage: Nirvana ABS vs AWS EBS
ClickBench reported 10.5× median speedup over tested gp3. Against io2, 77% of queries were within 1.5× runtime. Read the results and limitations.
Benchmarking Nirvana ABS against AWS EBS gp3 and io2 with ClickBench and FIO.
Updated 30 September 2026: corrected the io2 comparison in the graphics and metadata, separated the two gp3 median results, and clarified the pricing and test conditions.
What the benchmark reports
- Against the tested gp3 configuration: approximately 10.5× median speedup across the cold-read query suite, and 14.1× median speedup for its I/O-heavy subset.
- Against the tested io2 configuration: ABS query runtimes were within 1.5× of io2 on approximately 77% of queries and within 2× on approximately 98%.
- In the separate FIO tests: the report records approximately 20,000 IOPS per VM for an 80/20 read/write workload, with short bursts enabled.
- In the report's illustrative storage bill: ABS cost approximately 18% less than gp3 and 88% less than io2. This is a capacity-and-IOPS price example, not the bill for the ClickBench test.
These are Nirvana's reported results for specific configurations. They do not establish a performance advantage across every workload. See the benchmark summary and full report for the source methodology.
ClickBench cold-read results
The test ran 43 OLAP queries, Q0–Q42, against the hits_100m dataset. The OS page cache was dropped before each query. Each environment was benchmarked twice; the report analyzes cold-query timings from the second full run.
The reported configurations used 64 vCPUs, 512 GB of memory, 50 GB volumes and ext4. AWS used r7a.16xlarge instances. Nirvana matched the core count and architecture class with a lower CPU clock. Matching these specifications does not make the machines identical or eliminate every compute effect.
| Comparison | Reported result | What it measures |
|---|---|---|
| ABS versus gp3 | ~10.5× median speedup | Entire cold-read query suite |
| ABS versus gp3 | ~14.1× median speedup | I/O-heavy query subset |
| ABS versus io2 | ~77% within 1.5× runtime | Share of queries within the stated runtime factor |
| ABS versus io2 | ~98% within 2× runtime | Share of queries within the stated runtime factor |
The 77% figure is a share of queries, not a speedup. It does not mean ABS was 77% faster than io2, or faster on 77% of queries. A runtime within 1.5× can still be slower. The report says ABS was faster on a subset, but does not establish overall io2 parity.
The available report does not specify the gp3 throughput and IOPS provisioning or the io2 provisioned IOPS for this ClickBench run. Those settings matter to the comparison. Treat the speedups as results for the tested configurations, not a universal comparison with fully provisioned gp3 or io2. Cold-read results also do not predict warm-cache performance or vector-search latency.
FIO results and their limits
The separate FIO study used a 16-VM cluster with direct I/O through the production path: VM → Portworx CSI → iSCSI → ABS. Its baseline configuration used libaio, four jobs per VM, iodepth 32 and a 15-second runtime, with additional steady-state validation described in the report.
| Test | Reported observation per VM |
|---|---|
| 80/20 read/write mix | ~20,000 IOPS; ~1.3 ms average latency; short bursts enabled |
| 4K random writes | ~20,000 IOPS; ~3.5 ms latency |
| Sequential throughput | ~408–430 MB/s |
These are separate test observations, not one simultaneous operating point or a per-volume performance guarantee. Initial tiny-write tests at 2K and below performed worse; longer reruns improved them. Workload size, concurrency and the data path matter.
Price illustration with the assumptions visible
The original report's calculation uses 1,024 GB of billed capacity and a 10,000-IOPS requirement. It is separate from the 50 GB ClickBench volumes. The table reproduces the report's rates and arithmetic, rather than quoting a current offer.
| Storage | Calculation in the report | Monthly total |
|---|---|---|
| Nirvana ABS | 1,024 × $0.0935; IOPS included | $95.74 |
| AWS gp3 | 1,024 × $0.08 + 7,000 × $0.005 | $116.92 |
| AWS io2 | 1,024 × $0.125 + 10,000 × $0.065 | $778.00 |
Using the unrounded totals, ABS is 18.1% lower than gp3 and 87.7% lower than io2 in this illustration. Equivalently, io2 costs about 8.1 times as much. This comparison does not equalize throughput, latency, durability or features, and excludes compute and other infrastructure costs. The gp3 total includes no additional throughput charge beyond its baseline.
For a deployment decision, use the current Nirvana pricing and AWS EBS pricing for your region and settings, then test your own workload. AWS documents gp3 as provisioned performance without burst credits; it should not be described as a burst-credit product.
What to take from these results
The report supports a substantial advantage over the tested gp3 configuration for cold analytical reads. Against io2, the defensible claim is the measured runtime distribution above. Choose storage using the query mix, cache behavior, required latency and total deployment cost that apply to your application.
Read the benchmark · Explore Nirvana storage · Read the documentation
