LangChain Agent Infrastructure Benchmark: Nirvana ABS vs AWS EBS

ABS finishes 14-17% faster than io2 at every scale, at 31x less cost.

TL;DR

• ABS finishes 14-17% faster than AWS io2 at every scale (1K, 10K, 100K tasks)
• +19% more throughput, 7.8x faster raw disk I/O
• 31x cheaper ($118/mo vs $3,710/mo for io2-64k)
• io2-64k performs identically to io2-32k (instance-capped at ~40K IOPS)
• ABS wins every metric except Qdrant tail latency (p99). Over 100K tasks, throughput wins.
• All cold reads, 3 runs per scenario, fully reproducible

---

In our previous post we broke down what LangChain is, how the ReAct loop works, and where each step hits the cloud. The takeaway: every time an agent "acts," the data and storage layer (Qdrant, Redis, Postgres) reads and writes to the disk on your VM. That's the one layer where cloud storage performance directly affects how fast agents finish work.

This post is the results. We ran identical LangChain agent workloads across five storage platforms to answer one question: does the disk underneath actually make a difference?

It does.

14-17% faster task completion than AWS io2. +19% more throughput. 7.8x faster raw disk. 19x cheaper.

The workload

5M vectors across 50 Qdrant collections. 100 to 1,000 concurrent LangChain agents. Each task hits 3 services: Qdrant for vector search, Redis for caching, Postgres for checkpoints. 6 storage ops per task. All cold reads. Page cache dropped before every run.

The setup

Same workload, same hardware, five storage tiers. The only variable is the block storage.

Config	Instance	vCPU / RAM	Storage	Provisioned IOPS	Cost / mo
Nirvana ABS	n1-standard-4	4 / 16 GB DDR5	ABS 256 GB	20K baseline, 600K burst, included	$118
gp3-3k	m6i.xlarge	4 / 16 GB DDR4	gp3 256 GB	3,000	$147
gp3-16k	m6i.xlarge	4 / 16 GB DDR4	gp3 256 GB	16,000	$212
io2-32k	m6i.xlarge	4 / 16 GB DDR4	io2 256 GB	32,000	$2,238
io2-64k	m6i.xlarge	4 / 16 GB DDR4	io2 256 GB	64,000	$3,710

Held constant: instance class, 4 vCPU / 16 GB / 256 GB, same agent code and data on all five.

The only variable: the block storage. ABS includes 20,000 sustained IOPS in the base price. io2 is billed per provisioned IOPS.

Task completion at scale

ABS wins at every scale. The lead widens with load.

Scale	Tasks	Nirvana ABS	io2-32k	io2-64k	gp3-16k	gp3-3k	ABS
100x10	1,000	36s	43s	42s	43s	62s	14-16%
500x20	10,000	351s	420s	424s	430s	433s	16-17%
1000x100	100,000	58 min	69 min	69 min	68 min	71 min	16%

3 runs per scenario, all cold reads. Results reproduce across every run.

Task latency at 100K scale

Platform	Task p50	Task p95	Task p99
Nirvana ABS	326 ms	562 ms	725 ms
io2-32k	393 ms	612 ms	779 ms
io2-64k	396 ms	606 ms	771 ms
gp3-16k	390 ms	605 ms	760 ms
gp3-3k	404 ms	643 ms	809 ms

Per-service breakdown at 100K tasks

Each task chains 6 ops across 3 services. Here's where each platform wins and loses.

Service	Per task	ABS p99	io2-64k p99	ABS vs io2
Qdrant (vector search)	2 ops	301 ms	140 ms	io2 leads
Redis (cache)	2 ops	68.0 ms	70.4 ms	ABS 3% faster
Postgres (checkpoints)	2 ops	31.9 ms	37.7 ms	ABS 15% faster
Task p99 (compound)	6 ops	725 ms	771 ms	ABS 6% faster
Task completion	100K tasks	58 min	69 min	ABS 16% faster

io2 wins one metric: per-query Qdrant p99. ABS wins Redis p99, Postgres p99, task p99, and task completion.

Qdrant is 2 of 6 ops per task, not all of them. ABS wins Redis and Postgres. The 301ms Qdrant p99 is the worst 1% of queries. ABS Qdrant p50 is 66ms vs io2's 51ms, only 29% slower, not 117%. And ABS sustains 172 app IOPS vs io2's 145, 19% more tasks per second through the pipeline.

Over 100,000 tasks, throughput wins.

Cost

Platform	Monthly cost
Nirvana ABS	$118
gp3-3k	$147
gp3-16k	$212
io2-32k	$2,238
io2-64k	$3,710

ABS finishes faster than every AWS tier at a fraction of the cost. io2-64k performs identically to io2-32k on this instance class. The m6i.xlarge instance caps IOPS at ~40K regardless of what you provision.

Verdict

If you're running concurrent agent workloads: ABS. Fastest task completion at every scale we tested, at 19x less than io2.

Open source. Run it yourself.

We built this benchmark because nobody else did. LangChain has 138K stars and zero infrastructure benchmarks. Now there's one, and it's open source.

Open source. Run it yourself.

Full report: nirvanalabs.io/benchmark-results
Repo: github.com/nirvana-labs-examples/langchain-benchmarks

Terraform configs, Ansible playbooks, benchmark runner, raw JSON results.

Add your own storage config and compare.