Home › Topics › Ping, Traceroute & Friends › ICMP Politics

ICMP Politics: Why the Internet Lies to Your Probes


You ping a core router and it answers in 180 milliseconds — from thirty feet away. You traceroute to a cloud service and one hop in the middle is dropping half your probes. Yet the application riding that path is flawless. You ping a server that you can see is up, serving users, and get nothing but silence. New admins look at readings like these and conclude the network is on fire. The network is fine. The readings are politics.


Here is the uncomfortable truth every network veteran eventually internalizes: your probes are second-class citizens, and nobody owes them the truth. Ping and traceroute run on ICMP — the Internet Control Message Protocol. It is the network's side channel for service notes like "that address is unreachable" and "your packet died of old age." Forwarding user traffic is a router's job. Answering your probes is a favor. Devices grant the favor when convenient, ration it when busy, and refuse it on principle. Every one of those choices shows up in your terminal looking exactly like a network problem.


This guide is the decoder. You will learn the six mechanisms by which the internet routinely deceives probes — each with its telltale symptom. Then you will learn how to design monitoring that survives the deceit. Finally, you will learn how to run the ICMP policy on your own network like a good diplomat instead of a paranoid one.


In this guide

  • The side channel: what ICMP actually is
  • The six mechanisms of deceit
  • Designing checks that survive the politics
  • Diplomacy at home: your own ICMP policy
  • The decoder ring: symptom, cause, verification
  • Where to go next

The side channel: what ICMP actually is


Sixty seconds of protocol background, because the politics only make sense once you see what ICMP is for. TCP and UDP carry your actual data — web pages, files, voice. ICMP carries none of that. It is the network's administrative memo system: short messages between devices about the delivery process itself. "Destination unreachable." "Time to live exceeded." "Slow down." And the pair every admin knows by heart: echo request and echo reply — "are you there?" and "I am here" — which is all a ping is. Traceroute is cleverer. It deliberately provokes "time exceeded" memos from each router along a path. But it is reading the same memo system. (The full mechanics live in what ping actually tells you and reading traceroute.)


Here is the part that matters: delivering user traffic is what routers are built, benchmarked, and bought for. Answering memos is overhead. So the standards deliberately allow devices wide discretion in how — and whether — they respond to ICMP. Every network operator exercises that discretion differently. The result: the treatment your probe receives is a policy decision, not a physics measurement. Keep that sentence in your pocket. The next six sections are just that sentence wearing different hats.


The six mechanisms of deceit


1. Control-plane policing: the busy receptionist

A modern router is really two machines in one box. The data plane forwards traffic in specialized hardware at millions of packets per second without the CPU ever seeing them. That is the factory floor. The control plane is the router's modest CPU, which handles the thinking: routing protocols, management sessions, and, fatefully, generating ICMP replies. Your ping to the router's own address cannot be answered by the factory floor. It must climb up to the receptionist's desk, wait behind more important work, and be answered by hand. To protect that little CPU from being drowned, devices apply control-plane policing. These are strict limits on how much of this traffic the CPU will even accept.


The symptom: a router that answers pings slowly, jitterily, or intermittently — while forwarding user traffic through itself flawlessly. This is the single most misread symptom in networking: the scary 180 ms hop in a traceroute whose later hops are all 20 ms. The router is not slow at its job. It is slow at the favor. The receptionist being busy tells you nothing about the factory floor.


2. ICMP rate limiting: replies on a quota

Beyond protecting their CPUs, devices deliberately cap how many ICMP messages they will emit per second. This is a defense against being used as an amplifier in attacks, and against probe floods. The cap is often a plain token bucket. The first N requests per second get replies. The rest get silence. Your probes above the quota are not lost in any meaningful sense. They are declined.


The symptom: patterned loss to a device that is demonstrably fine. The loss may sit at a suspiciously round plateau (say, right at 50% when you probe at twice the quota). Or it may kick in only when you probe fast and vanish when you slow down. Or it may affect the fifth ping of every burst. Organic network loss is ragged and random. Quota loss has geometry. If the loss changes when you change your probe rate, you are measuring the quota, not the network.


3. Firewalls: silence is not absence

Many firewalls drop echo requests entirely — no reply, no error, just a shrug. At many organizations that is deliberate, written policy. The policy is: don't advertise what exists, don't answer strangers, and accept that ping goes dark. Whole enterprises run this way permanently, with servers that have never answered a ping in their lives and never will. Security scanners even flag ping-answering hosts as findings, so the practice is self-reinforcing.


The symptom: total, permanent silence from a host that is visibly alive and serving. The lesson is one of the most important in this whole topic: silence proves nothing. "No ping reply" means "no ping reply" — it does not mean down. Treating unpingable as down will flood you with false alarms in any environment with a security team worth its salary.


4. Asymmetric treatment: polite in one direction only

ICMP policy is applied per direction, and the two directions of a round trip can be governed by different administrations with different opinions. A firewall may allow your echo request out but drop the echo reply coming back. A provider may deprioritize ICMP only on its inbound edge. Your request and its reply may even travel physically different routes. Asymmetric routing is normal on the internet. There may be politics on one route and clean passage on the other.


The symptom: a mystery with a hidden half. Ping reports one number for the whole round trip. So a delay or drop that lives purely in the return direction is indistinguishable from one on the way out. You blame the destination. The destination's operators, probing from their side, see nothing wrong and blame you. Both measurements are honest. Each is blind to the other half of the trip.


5. QoS classing: the first-class-probe illusion

QoS (Quality of Service) machinery sorts traffic into classes and queues them differently under load — voice first, bulk transfers last. The question nobody asks: which queue do your probes ride in? Almost never the one carrying the traffic they claim to represent. The deceit runs in both directions. Your probes can look better than reality. Tiny ICMP packets slip through queues that are mauling the 1,500-byte data packets your users actually send. So the graph is green while file transfers crawl. And they can look worse. Plenty of networks shove ICMP into the scavenger class that absorbs congestion first. So your graph screams during busy hour while the prioritized voice and business traffic glides through untouched.


The symptom: probe measurements that diverge from user experience under load, in either direction — green graphs with complaining users, or red graphs with happy ones. Divergence under load specifically is the tell, because queues only sort when there is a line.


6. Middleboxes: the doorman who answers for everyone

Between you and many services stands equipment that answers on the service's behalf. Examples include load balancers holding the public address for a pool of servers, proxies, VPN concentrators, anycast nodes. Ping the service's address and the thing that replies is the doorman, not the building. The load balancer will cheerfully answer echo requests while every backend server behind it is down. That is because the load balancer is, in fact, up. It is just telling you about itself.


The symptom: the cruelest one — a dead service that pings beautifully. Users cannot log in. Your reachability check is green. The address answers in 8 ms with zero loss. That is because the front door is in excellent health and the check never asked about anything behind it.


Designing checks that survive the politics


You cannot repeal the politics. You can design monitoring that stops being fooled by it. Four habits do most of the work:


  • Probe the service, not just the address. The doorman problem and the firewall problem share one cure: ask the question you actually care about. Try a TCP connection to the real port, a real DNS query, a real login-page fetch. These traverse the same policy as user traffic and get answered by the thing users depend on. Ping tells you an address answers. Only a service check tells you the service works. That is the distinction ping-up versus service-up exists to hammer home. The practical toolbox is next door in Beyond ICMP: TCP and UDP reachability checks.
  • Treat intermediate-hop readings as gossip. A middle hop's latency or loss in a traceroute is the receptionist's mood, not the factory's throughput. Judge a path by its destination. End-to-end numbers describe what traffic experiences. Per-hop numbers describe how each router feels about answering memos. Believe a middle hop only when the trouble starts there and continues through every hop after it.
  • Test from more than one vantage point. Asymmetry and one-sided policy are invisible from a single seat. A probe at the far site looking back, or a check from a second network, turns "the internet is weird" into "the outbound direction through provider A is weird." Placement strategy is its own craft — see designing synthetic probes.
  • When you must ping a device, baseline its politics. Some things — routers, switches, printers — you will monitor by ping anyway. That is fine once you learn that device's normal. Watch it for a week. For example, this router answers in 2 ms except during its nightly config backup. That firewall never answers more than five per second. The old switch drops every ping when its CPU touches 80%. Then alert on change from its normal, not on absolute numbers. A device whose politics you have baselined is a perfectly good canary. A device you started monitoring an hour ago is a slot machine.

Tip: before escalating anything based on a scary middle-hop reading, check the final hop. If the destination's latency and loss are clean, the middle hop's drama is almost certainly control-plane theater — interesting, harmless, and not your outage.


Diplomacy at home: your own ICMP policy


Now turn the telescope around: on your own network, you are the government. Other people's probes — and more importantly, the internet's legitimate service memos — are subject to your policy. Two principles keep you out of the classic traps.


Allow sensible ICMP inside your network

Internally, ping is a diagnostic gift: let your admin subnets and monitoring systems ping freely, and let troubleshooting stay easy. The blanket "block all ICMP everywhere" instinct feels like security and mostly manufactures mysteries — because ICMP is not just ping. Some of those memos are load-bearing, and the classic self-inflicted wound proves it:


The path MTU black hole. Every link has a maximum packet size (the MTU — maximum transmission unit). A router may need to forward a packet too big for the next link. If that packet is flagged "don't fragment" (as modern TCP traffic is), the router drops it and sends back an ICMP memo. The memo says "fragmentation needed" — effectively, send smaller ones. The sender shrinks its packets and life continues. This dance is path MTU discovery. It is happening constantly on your behalf. Now imagine blanket-blocking ICMP at your edge. In that case, the memo never arrives. The sender keeps launching full-size packets into a void. You get the strangest ticket in networking — connections that start fine and then hang. Small packets fit; big ones vanish. Web pages half-load. Emails send until someone attaches a file. Everything works except through the VPN, whose tunnel overhead shrank the MTU. If you have ever met a network where "it connects but big transfers stall," you have met a path MTU black hole. Somewhere upstream a firewall admin was very proud of blocking all ICMP.


You can hunt this one yourself by sending pings of a chosen size with fragmentation forbidden, and watching where they start to die:


# Windows: 1472-byte payload, don't fragment
# (1472 + 28 bytes of headers = a full 1500-byte packet)
ping -f -l 1472 server.example.com

# Linux equivalent
ping -M do -s 1472 server.example.com

# A healthy path answers. A too-big packet gets:
#   "Packet needs to be fragmented but DF set."
# Step the size down until replies return; that reveals
# the path's real MTU. Silence at every size while normal
# ping works = the "fragmentation needed" memos are being
# blocked somewhere. That is the black hole.

A workable internal policy, in one list — permit these and you keep the diagnostics and the plumbing:


  • Echo request and reply — at minimum from management and monitoring subnets; ping is how your own tools breathe.
  • Destination unreachable — especially the "fragmentation needed" variety; blocking it builds the black hole above.
  • Time exceeded — what traceroute lives on; block it and you blind your own path diagnostics.

Warning: in IPv6 the stakes jump from "annoying" to "broken." Routers do not fragment IPv6 packets at all. So the "packet too big" memo is the only mechanism that makes mixed-MTU paths work. And neighbor discovery, IPv6's replacement for ARP, is also carried in ICMPv6. Blanket-block ICMPv6 and you have not hardened your network; you have unplugged it politely.


At the edge: rate-limit, don't blackhole

Facing the internet, you have legitimate reasons not to answer every stranger's probe: reconnaissance resistance, amplification hygiene, CPU protection. The graceful implementation is the one the grown-up networks use: rate-limit ICMP instead of blackholing it. Answer a civilized trickle of echo requests, keep emitting the load-bearing memos, and cap the volume so nobody can weaponize you. You get the security properties without manufacturing black holes for your own users. And your provider can still ping your edge when you call them at 2 AM insisting the circuit is down. That is exactly when you want your edge to be pingable by someone.


The decoder ring: symptom, cause, verification


Pin this table up. When a probe reading looks insane, find the symptom, suspect the politics, and run the verification before escalating anything:


SymptomLikely political causeHow to verify
One middle hop shows huge latency; later hops are fineControl-plane policing — the router deprioritizes its own repliesJudge the destination's numbers, not the hop's; confirm end-to-end latency is clean
Loss to a device sits at a neat plateau, or changes when you change probe rateICMP rate limiting — you are measuring a quotaSlow your probe rate and watch loss vanish; check the device's own health over SNMP; test loss through it versus to it
Host never answers ping but visibly serves usersFirewall drops echo by policyTCP-check the real service port; ask whoever owns the firewall — the block is probably intentional
You see loss; the far side, probing back, sees noneAsymmetric routing or one-direction ICMP policyProbe from both ends at once; compare paths with traceroute from each side
Graphs green under load while users suffer (or red while users are happy)QoS classing — probes ride a different queue than real trafficCompare with TCP handshake timing to the real port during busy hour; watch for divergence only under load
Address pings perfectly; service is deadA middlebox answers on the service's behalfApplication-layer check behind the front door — real query, real login, real response
Sessions connect, then hang on big transfers; small stuff finePath MTU black hole — someone blocked "fragmentation needed"Don't-fragment pings, stepping the size down; fix the ICMP block, not the application

Where to go next


Once you accept that probe treatment is policy, the readings stop being lies and become a second data stream. Ping tells you how a device chooses to answer. Baselined and watched for change, that is still useful intelligence. The discipline is refusing to confuse the favor with the physics. Judge services at the service layer, judge paths end to end, and treat middle hops as gossip. And keep your own network's memos flowing so you never black-hole yourself.


From here, a natural next read is Beyond ICMP: TCP and UDP reachability checks — the practical toolbox for probing above the politics. Then read reading traceroute like a professional, which applies today's skepticism hop by hop. For the long-haul view of what honest probing looks like over time, see smokeping-style latency graphing. For the full toolbox, the Ping, Traceroute & Friends pillar has the map.


Put this into practice with SysPeep Monitor

You can't change the internet's politics, but you can stop guessing about your own equipment. SysPeep Monitor watches your devices where the truth lives. It collects health and performance data over SNMP, WMI, and SSH from one Windows machine. Its dependency-aware alert rules help you hear real failures instead of noise. When something goes quiet, you'll know whether it's diplomacy or an outage.

Download Free TrialLearn About SysPeep