Benchmarks & Cloud Financial ROI

TriCache is engineered and benchmarked against high-concurrency microservice and edge workloads.

---

In-Memory Warm Read Throughput

Measured via native Node.js benchmarks (pnpm bench) with V8 garbage collection exposure on an 8-core machine:

Cache Engine	Architecture	Throughput (ops/sec)	Latency (ns/op)	Allocation Overhead
TriCache (L1 RAM)	W-TinyLFU + Segmented LRU	2,810,000 ops/s	356 ns	0 bytes (zero-alloc)
lru-cache v11	Doubly-linked list LRU	2,150,000 ops/s	465 ns	Minimal
@neshca/cache-handler	In-Memory + Redis Adapter	820,000 ops/s	1,219 ns	Medium (wrapper overhead)
keyv (in-memory)	Map wrapper	1,400,000 ops/s	714 ns	Low
Single-Tier Redis	Local VPC Network Hop	8,500 ops/s	1,200,000 ns (1.2 ms)	Network TCP socket

Warm in-memory reads in TriCache run at 2.81 million ops/sec, bypassing V8 deserialization by referencing hot cached JS primitives.

---

Tier Latency Hierarchy

Latency Hierarchy Overview

┌─────────────────────────────────────────────────────────────┐
│  Tier 1: L1 RAM (W-TinyLFU)                ~350 ns          │
│  Tier 1.5: /dev/shm (tmpfs off-heap RAM)   ~20 µs           │
│  Tier 2: L2 Redis (Network VPC)            ~1,200 µs (1.2ms)│
│  Tier 3: Database / Upstream API           ~50,000 µs (50ms)│
└─────────────────────────────────────────────────────────────┘

---

Cloud ROI & Adversarial Simulation

To evaluate real-world infrastructure impact without theoretical bias, TriCache includes a deterministic cloud simulation based on an adversarial Zipfian workload ($N=10,000$ keys, skew $s=0.99$, 100,000 requests, 15% novel cold-tail misses, 5% writes, 2KB payloads, seeded via Mulberry32 0xDEADBEEF).

Run the simulation locally:

pnpm bench:roi

1. Operational Throughput & Tier Distribution

Throughput & Tier Distribution Data

+--------------------------+---------------------+---------------------+
| Metric                   | Single-Tier Redis   | TriCache 3-Tier     |
+--------------------------+---------------------+---------------------+
| L1 RAM Hits              | 0 (0.0%)            | 48,455 (48.5%)      |
| L1.5 /dev/shm Disk Hits  | 0 (0.0%)            | 14,004 (14.0%)      |
| L2 Redis Hits            | 72,901 (72.9%)      | 10,442 (10.4%)      |
| DB Misses (fetchFn)      | 22,082 (22.1%)      | 22,082 (22.1%)      |
| Total Redis Commands     | 122,082             | 59,623              |
| Redis Command Reduction  | Baseline (0.0%)     | -51.2%              |
+--------------------------+---------------------+---------------------+

2. Latency Percentile Compression

Percentile Compression Data

+--------------------------+---------------------+---------------------+
| Percentile               | Single-Tier Redis   | TriCache 3-Tier     |
+--------------------------+---------------------+---------------------+
| p50 (Median)             | 1.50 ms             | 0.08 ms (18.7x)     |
| p95                      | 51.50 ms            | 50.00 ms            |
| p99                      | 51.50 ms            | 50.00 ms            |
+--------------------------+---------------------+---------------------+

3. AWS Infrastructure Financial Savings (100M req/day fleet)

Financial Savings Breakdown

+-------------------------------------+------------------+------------------+
| Cloud Infrastructure Line Item      | Single-Tier      | TriCache 3-Tier  |
+-------------------------------------+------------------+------------------+
| AWS ElastiCache Cluster Instance    | 3x r6g.xlarge    | 2x r6g.large     |
| ElastiCache Annual Instance Cost    | $10,459.44       | $3,486.48        |
| Monthly VPC Inter-AZ Bandwidth      | 5,722 GB/mo      | 2,148 GB/mo      |
| Annual AWS VPC Data Transfer Cost   | $686.65          | $257.77          |
+-------------------------------------+------------------+------------------+
| TOTAL ANNUAL CLOUD INFRASTRUCTURE   | $11,146.09       | $3,744.25        |
+-------------------------------------+------------------+------------------+
| NET ANNUAL SAVINGS WITH TRICACHE    | ->               | $7,401.83        |
+-------------------------------------+------------------+------------------+

Executive Takeaway

TriCache absorbs 51.2% of Redis command volume directly in local RAM and /dev/shm, dropping median latency by 18.7× and saving $7,401.83/year per cluster while protecting upstream databases from stampedes.

---

Granular Micro-Benchmarks

Data collected via pnpm bench under Node.js $\ge 22$:

1. Bloom Filter Probe Cost

The WASM Bloom filter is $O(k=7)$ per probe. Definite misses abort without accessing the V8 Map:

Operation	Throughput	Latency	Notes
Definite miss (never set)	5,360,000 ops/s	187 ns	7 hash rounds $\to$ bit check $\to$ return null
Hit path (confirmed present)	3,220,000 ops/s	310 ns	7 hash rounds $\to$ Map.get $\to$ return cached entry

2. Serialization Throughput (msgpackr 2.1.0)

Payload Size	Throughput	Latency
128 B	827,700 ops/s	1.21 µs
256 B	706,200 ops/s	1.42 µs
512 B	625,800 ops/s	1.60 µs
1,024 B (1 KB)	472,300 ops/s	2.12 µs
4,096 B (4 KB)	228,600 ops/s	4.37 µs

3. Cryptographic Ciphers & Obfuscation (L2 Redis & Disk)

Mode	Payload Size	Encrypt Latency	Decrypt Latency	Throughput
AES-256-GCM	64 B	7.12 µs	6.43 µs	155,500 ops/s
AES-256-GCM	4 KB	17.12 µs	20.84 µs	58,400 ops/s
AES-128-GCM	64 B	6.72 µs	5.78 µs	173,000 ops/s
AES-128-CTR	64 B	5.32 µs	5.08 µs	196,900 ops/s
XOR Obfuscation	64 B (32-bit CPU path)	0.41 µs (412 ns)	0.47 µs (476 ns)	2,430,000 ops/s

4. L1 Memory Iteration (500 live entries)

Method	Throughput	Latency	Notes
cache.keys()	26,600 ops/s	37.53 µs	No [key, entry] tuple allocation
cache.values()	35,500 ops/s	28.19 µs	Direct yield* delegation
cache.entries()	24,000 ops/s	41.73 µs	Stripped namespace keys
cache.scan(fn)	210,000 ops/s	4.76 µs	Zero-alloc direct index scan