Skip to content
Notifications
Clear all

Has anyone benchmarked the latency added for East Coast users hitting EU DCs?

2 Posts
2 Users
0 Reactions
0 Views
(@hiroshim)
Reputable Member
Joined: 1 week ago
Posts: 188
Topic starter   [#10333]

I am conducting a preliminary analysis for a global financial application architecture where we are considering consolidating our European user base into a single Prisma Access hub located in Frankfurt. However, a significant portion of our internal users and data sources remain on the US East Coast (specifically, Northern Virginia AWS us-east-1). I am seeking to quantify the inherent latency penalty introduced by Prisma Access's traffic hairpinning in this specific corridor.

The concern is straightforward: a user in London, authenticated via the Frankfurt Prisma Access hub, requesting data from an origin server in us-east-1, would incur the following network path:
`User (London) -> Frankfurt Prisma Access POP -> Origin Server (us-east-1)`.
This is compared to a direct, optimized path like:
`User (London) -> us-east-1`.

My initial back-of-the-envelope calculation, based on public internet light-speed latency, suggests a baseline of approximately 70-80ms for a direct London-to-Northern Virginia path. Routing through Frankfurt could logically add the latency of LondonFrankfurt (approx. 15-20ms) plus Frankfurtus-east-1 (approx. 85-95ms), potentially doubling the RTT.

I am looking for empirical, production-level data from the community. Has anyone performed structured benchmarking or has ongoing monitoring that captures this specific scenario? Vendor-provided "typical latency" figures are insufficient for a performance-sensitive design.

Ideal data points would include:

* **Test Methodology:** e.g., Synthetic ICMP pings, HTTP request/response cycles, or TCP handshake measurements.
* **Source Geography:** Specific city or region of the user/source.
* **Prisma Access Hub Location:** e.g., "Germany" or "Frankfurt".
* **Destination:** Specific cloud region and service (e.g., AWS EC2 in us-east-1, Azure SQL in East US 2).
* **Measured Latency:** Both to the Prisma Access POP (first hop) and to the final destination, highlighting the added latency.
* **Time of Day & Sample Size:** To account for periodic congestion.

For example, a useful benchmark output might look like this (from a hypothetical monitoring script):

```bash
# Direct path from London EC2 instance to us-east-1 RDS
Target: database.us-east-1.rds.amazonaws.com
Min: 72.3ms Avg: 74.1ms Max: 81.5ms Loss: 0%

# Path from same London EC2, via Prisma Access (Frankfurt hub) to same RDS
Target: database.us-east-1.rds.amazonaws.com (via Prisma Tunnel)
Min: 154.8ms Avg: 162.4ms Max: 178.9ms Loss: 0%
Latency Added by Hairpinning: ~88.3ms
```

This data is critical for a trade-off analysis between security posture (single, consolidated inspection point) and application performance. The additional latency directly impacts user experience for real-time applications and influences whether we need to consider a multi-hub architecture or accept the performance tax. Any shared experiences, including the impact of Prisma Access's own network optimizations (like dedicated egress), would be highly valuable.



   
Quote
(@integrations_jane)
Reputable Member
Joined: 3 months ago
Posts: 172
 

You're in the right ballpark with those numbers, but the real-world penalty can be much worse than just summed latencies. The hairpin adds a processing penalty at the Frankfurt POP that's variable based on load and traffic inspection profiles. I've seen it add a consistent 8-
15ms on top of the pure network hop in that corridor.

Also, don't forget that the `User -> Frankfurt` leg for someone in London isn't necessarily a clean internet path. It's the corporate VPN tunnel into Prisma Access, which might itself be taking a suboptimal route before it even hits the hub. Your 70-80ms direct baseline is ideal; in practice, I'd start with 95ms as a floor and expect the hairpinned scenario to push 150ms or more for TCP handshakes. For a financial app, that extra 60-80ms on every synchronous request adds up painfully.


APIs are not magic.


   
ReplyQuote