The Physics of Jitter

In high-frequency trading (HFT), we are taught that speed is king. But in the modern institutional environment, speed is a commodity. The true battleground has shifted from speed to determinism.

The Average Latency Trap

Most infrastructure providers (AWS KMS, Fireblocks, etc.) market their “average latency.” But in a competitive order book, your average latency doesn’t matter. What matters is your tail latency—the 1-in-1,000 event where a 150ms spike causes you to miss an insertion window or results in an adverse price fill.

We call this the Jitter Tax.

The Mathematical Model of Jitter Tax

We define the Jitter Tax ($T_j$) as the expected value of alpha loss over time:

$T_j = \sum_{i=1}^{n} V_i \cdot s \cdot \max(L_i - L_{base},\ 0)$

Where:

• $V_i$ is the notional volume of trade $i$.
• $s$ is the slippage rate (P&L loss per unit of excess delay, in $/ms/unit volume).
• $L_i$ is the observed signing latency for trade $i$.
• $L_{base}$ is the deterministic lower bound for the signing operation (hardware floor).

The $\max(\cdot, 0)$ term ensures only latency above the baseline contributes to the tax. Any realization where $L_i > L_{base}$ represents a trade that fired late—either missing the insertion window entirely or taking an adverse fill.

Latency Characterization

The following figures reflect the latency profile of each signing method. AWS KMS and CloudHSM numbers are based on AWS-published performance characteristics and community benchmarks for same-region clients. Sentinel figures are from Rust benchmark measurements of ECDSA signing (secp256k1, k256 crate) running within the Nitro Enclave process, communicating over VSOCK—no external network hop.

Provider	Median Latency	P99.9 Latency	Jitter Ratio (P99.9/Median)	Baseline
AWS KMS (same-region API)	~130ms	~400ms	~3x	Public API, TLS + HSM queue
AWS CloudHSM (same-VPC)	~2ms	~15ms	~7.5x	Private HSM, still network-bound
Sentinel (local enclave, VSOCK)	~42µs	<100µs	<2.5x	In-process signing, no network hop

Important baseline notes:

• AWS KMS median latency varies by region and request concurrency. AWS publishes a p50 target of ~100–130ms for Sign API calls in us-east-1 under typical load; the ~130ms figure above reflects this. The P99.9 figure reflects behavior under sustained concurrency, not single-shot measurements.
• AWS CloudHSM is significantly faster than KMS because it bypasses the managed API surface, but it still routes traffic over the VPC network.
• The Sentinel 42µs figure is the measured median for raw secp256k1 ECDSA signing within the enclave process (k256 Rust crate, SHA-256 prehash). This measures the signing operation itself. Total end-to-end latency including VSOCK transport adds a small increment (typically single-digit microseconds on the same host).
• The Jitter Ratio for Sentinel is low because the signing operation runs in a pinned, isolated vCPU context with no kernel I/O or network stack traversal.

Why Tail Latency Matters More Than Median

The Jitter Tax is not a function of median latency—it is a function of the distribution’s right tail. A system with 130ms median and 400ms P99.9 creates a situation where roughly 1 in 1,000 orders fires 270ms late. At institutional volumes (thousands of order events per second), this is not an edge case; it is a recurring drag.

The critical insight is that tail latency in network-bound signing correlates with peak market activity—the exact moments when execution quality matters most. Signing queues at KMS or CloudHSM back up precisely when transaction volume is highest, amplifying the Jitter Tax when alpha is richest.

Conclusion

The Jitter Tax is not eliminated by faster average signing. It is eliminated by removing the network hop entirely. Moving signing logic into a local Nitro Enclave—communicating over VSOCK rather than HTTPS—shifts the bottleneck from network and HSM queue latency to raw CPU arithmetic. That is a qualitatively different operating regime, not just a quantitative improvement.

---

This report is part of the ZeroCopy Systems Research Series. Benchmark methodology and raw results are available on request.