How to Test China Routing on a VPS: Looking Glass, TCP MTR, and Packet Size Checks
A VPS can answer a ping from Shanghai in 42 milliseconds with zero packet loss and still stall the moment a mainland China user loads an HTTPS page or pulls a Docker image. Default network checks miss that failure because they test a single direction with tiny ICMP packets:
- Outbound tests from a laptop in mainland China only reveal the forward path to the server. Web responses, file downloads, and video streams travel on the return path from the VPS back to mainland China, which often crosses different carrier links.
- Standard
pingand defaultmtrsend ICMP Echo probes. Policy-based routing and carrier quality-of-service (QoS) rules can route or police TCP traffic on port 443 differently from ICMP. - A 64-byte probe slips through congested queues and hidden tunnel encapsulations that drop or fragment 1,400-byte data segments.
Verifying a China-optimized VPS means testing the return path across all three state carriers before your refund window closes, using both TCP probes and full-sized packets.
Check the return path first with a Looking Glass
Internet routing is asymmetric. Each network along a path makes an independent forwarding decision based on its own BGP table and commercial transit agreements. A request from a China Telecom user in Suzhou might enter Tokyo over a clean path, while the server’s reply returns through a congested public peering exchange or a standard backbone.
For most services, the return path to mainland China carries the bulk of the bytes. Inbound packets from the client are mostly small TCP acknowledgments and HTTP requests. Outbound packets from the VPS carry TLS certificate chains, HTML, API payloads, and media segments. Even when you only care about uploads, TCP acknowledgments returning from the server still govern how fast the client can send.
Before ordering a server, open the provider’s Looking Glass. Riven Cloud runs public Looking Glass nodes in Tokyo and Singapore connected to the same production network as customer VPS instances. Run a traceroute from the Looking Glass toward a real IP address in the mainland province and carrier your users connect from, rather than a generic resolver like 8.8.8.8.
In a three-network setup covering China Telecom, China Unicom, and China Mobile, check the border handoff where traffic leaves the international upstream and enters the Chinese carrier’s autonomous system (ASN):
| Carrier | Premium return path | Standard congested backbone | What to look for on the return trace |
|---|---|---|---|
| China Telecom | CTGNet (formerly China Telecom CN2 GIA), AS23764 and AS4809 | ChinaNet (163), AS4134 | AS23764 handing off to AS4809 (59.43.x.x hops) into the mainland gateway |
| China Unicom | China Unicom Premium (CUP), AS10099 and AS9929 | Unicom 169, AS4837 | AS10099 pairing with AS9929 (218.105.x.x or 210.51.x.x) before provincial access hops |
| China Mobile | China Mobile International N2 (CMIN2), AS58807 | CMI standard (AS58453) to CMNET (AS9808) | AS58807 on the international leg before handing off to domestic AS9808 |
Our guide on how to read China’s carrier AS numbers explains how these networks interconnect and why hop prefixes alone are hints that should be backed by ASN lookups. If a provider claims three-network optimization but the Looking Glass trace toward a China Telecom IP enters AS4134 (202.97.x.x) at the international border, you are looking at standard ChinaNet transit or partial CN2 GT routing. Our list of six questions to ask about CTGNet and CN2 GIA claims covers the commercial side of that check.
A web-based Looking Glass has one limit: it only runs the probe type the web interface exposes, usually default ICMP or UDP traceroute with small packets. Once the Looking Glass confirms the BGP and ASN structure, run protocol and packet-size checks from a shell on the VPS itself.
Test TCP routing alongside ICMP
Default mtr on Linux sends ICMP Echo Request packets. Relying on ICMP alone leaves two blind spots when evaluating routing to mainland China.
First, an upstream provider or reseller with limited premium capacity can configure policy-based routing (PBR) at the edge. By matching the protocol field in the IP header, a router can forward ICMP traffic over a low-latency CTGNet or AS9929 circuit while steering bulk TCP traffic on ports 80 and 443 onto cheaper transit such as AS4134 or a third-party regional carrier. To a casual buyer running ping or a basic mtr, the server looks fast even while real HTTPS connections crawl.
Second, carrier core routers treat control-plane ICMP responses differently from forwarded TCP packets. When a router decrements a probe’s time-to-live (TTL) to zero, its route processor generates an ICMP Time Exceeded message. Most carrier routers apply control-plane policing (CoPP) to those replies. Conversely, firewall filters or middleboxes at cross-border gateways can drop or deprioritize TCP SYN packets on specific ports while letting ICMP Echo pass untouched.
From your KVM VPS, install mtr and compare a 100-cycle ICMP report against a TCP report targeting port 443 toward the same mainland destination IP:
# Baseline ICMP return path with ASN lookup
mtr -r -c 100 -z 203.0.113.10
# TCP return path on port 443 (requires root or cap_net_raw)
sudo mtr -r -c 100 -z -T -P 443 203.0.113.10
If you prefer traditional traceroute, run the equivalent TCP SYN trace:
sudo traceroute -T -p 443 -A 203.0.113.10
When you compare the AS column and hop IP addresses across both outputs, both traces should enter the same upstream and cross into the mainland carrier through the same border AS (AS23764 / AS4809, AS10099 / AS9929, or AS58807). Small differences in load-balanced router interfaces within the same AS are normal because Equal-Cost Multi-Path (ECMP) hashing includes the TCP port numbers alongside the IP addresses. If the ICMP trace enters China via AS4809 (59.43.x.x) while the -T -P 443 trace detours through AS4134 (202.97.x.x) or an unrelated transit provider, the host is splitting routes by protocol.
Watch how packet loss behaves down the column as well. If hop 5 shows 40% loss in the Loss% column but hops 6 through 12 show 0.0% loss, hop 5 is only rate-limiting ICMP Time Exceeded replies. Real path loss starts at a specific hop, often the international border handoff, and continues all the way to the final destination hop.
Compare small probes against MTU-sized packets
Passing a default TCP MTR test proves the packets follow the right path, yet throughput can still collapse under load. By default, mtr sends 64-byte packets and standard ping sends 84 bytes including the IP and ICMP headers, whereas a real HTTPS session immediately sends its TLS certificate chain and encrypted application data in 1,400- to 1,500-byte TCP segments. If you have ever wondered why HTTPS handshakes are slow from mainland China, packet loss on those initial multi-kilobyte flights is a common culprit.
Testing with both small and large packets exposes two operational problems. On an oversold cross-border link, a 64-byte probe fits into shallow router queues or token-bucket policers, while a burst of 1,400-byte packets fills the buffer and triggers tail drop or random early detection (RED). Some traffic shapers also penalize full-sized frames while letting small interactive packets pass to keep latency figures low.
Large packets also reveal hidden tunnels. Some budget hosts lack direct BGP sessions with premium Chinese carriers at the local data center and haul traffic across GRE, IPsec, or WireGuard tunnels to another facility. Tunnel headers consume 24 to 60 bytes of the 1,500-byte Ethernet maximum transmission unit (MTU). When Path MTU Discovery (PMTUD) fails because a firewall drops ICMP Fragmentation Needed messages, large TCP segments with the Don’t Fragment (DF) bit set vanish silently, causing connections to hang right after the TCP handshake.
Test packet size sensitivity by running two MTR reports back to back with the -s (packet size) flag:
# Small 64-byte probe baseline
sudo mtr -r -c 100 -z -T -P 443 -s 64 203.0.113.10
# Large 1400-byte payload test
sudo mtr -r -c 100 -z -T -P 443 -s 1400 203.0.113.10
A clean premium route keeps final-hop Loss% near zero on both 64-byte and 1,400-byte probes. If 64-byte probes show 0% loss while 1,400-byte probes show 5% to 15% loss from the carrier border onward, the cross-border circuit is congested or shaped. Check the StDev (standard deviation / jitter) column too: a 1 to 3 ms increase in serialization delay for larger packets is normal physics, whereas a jump from 1.5 ms to 25 ms on 1,400-byte packets points to queuing delay at a bottleneck interface.
If the -s 1400 trace stops responding altogether after a specific upstream hop, test the exact path MTU using ping with the Don’t Fragment flag enabled:
# 1472 bytes payload + 8 bytes ICMP + 20 bytes IPv4 = 1500 bytes total frame
ping -M do -s 1472 -c 20 203.0.113.10
On a native 1,500-byte path, ping -M do -s 1472 succeeds toward targets that accept full-sized frames (some mainland residential PPPoE connections cap MTU at 1,492 bytes, which requires -s 1464). If probes fail above 1,400 bytes right at the VPS provider’s upstream gateway, the traffic is riding inside an unannounced tunnel.
Run your checks during China evening peak across all three carriers
At 10:00 China Standard Time (CST, UTC+8), even standard ChinaNet (AS4134) and Unicom 169 (AS4837) links often run with low loss and stable latency. Congestion on international gateways builds between 20:00 and 23:00 CST, when mainland residential usage peaks. That three-hour window is what you pay to avoid when buying CTGNet, CUP, and CMIN2 transit. Providers that oversell their premium commit or rely on dynamic BGP routing that fails over to cheap transit under load only reveal themselves during evening peak.
Run your daytime baseline and your 20:30 to 22:30 CST peak test against at least one target IP on China Telecom, China Unicom, and China Mobile in the provinces you serve. A host that routes China Telecom over AS4809 while handing China Mobile to congested regional peering only solves a third of the problem.
A pre-deployment verification checklist
When evaluating a new Tokyo or Singapore VPS for mainland China traffic, run through these five steps in order:
- Run a traceroute and BGP lookup from the provider’s Looking Glass to mainland China Telecom, China Unicom, and China Mobile IPs to confirm the return path uses
AS23764/AS4809,AS10099/AS9929, andAS58807. - Provision the VPS and run
sudo mtr -r -c 100 -z -T -P 443toward your target IPs to verify that TCP traffic follows the same premium AS path as ICMP. - Repeat the MTR test with
-s 1400and compare final-hop packet loss andStDevjitter against a-s 64baseline. - Run
ping -M do -s 1472(or-s 1464for PPPoE endpoints) to verify that the path supports standard MTU frames without hidden tunnel fragmentation. - Re-run the TCP and 1,400-byte tests between 20:00 and 23:00 CST to confirm the route does not degrade or switch AS paths during evening peak.
You can run the first step right now on Riven Cloud’s Tokyo Looking Glass and Singapore Looking Glass, or review our upstream ASN architecture on the network page. Every Riven Cloud KVM VPS plan, starting with the Premium plan at $59.95/month, includes identical three-network routing over CTGNet, CUP, and CMIN2, backed by a 14-day refund window (for servers under 10 GB of transfer) so you can run your own TCP and large-packet MTR benchmarks during peak hours before committing.
Share