“Echo works.” I have heard that sentence used to close a ticket, sign off a go-live, and end an argument with a vendor. It should not do any of those things.
Strictly speaking, C-ECHO is a connectivity test. The standard defines Verification as exactly that: confirming that two DICOM application entities can communicate. What it isn’t is a test of whether those two systems can do the work you’re connecting them for. A successful C-ECHO proves that an association was accepted and that the far side answered a Verification request. That’s a real result, and it rules out a whole class of problems. But it says nothing about whether a single image will ever cross that connection, and the problems that actually stop images are almost all in the part echo doesn’t touch.
This is what echo tests, what it skips, and the sequence I use instead.
What C-ECHO actually does#
The Verification SOP Class (PS3.4, Annex A) is the simplest service in DICOM. The SCU opens an association proposing a Verification presentation context, which in a conformant SCU includes implicit VR little endian (the one transfer syntax every conformant implementation has to support, PS3.5 Section 10.1), sends a C-ECHO request with no data set, and the SCP responds with a status. Success is 0000.
sequenceDiagram
participant SCU as Your tool
participant SCP as Target AE
SCU->>SCP: TCP connect to host:port
SCU->>SCP: A-ASSOCIATE-RQ
(calling AE, called AE, Verification SOP Class)
SCP-->>SCU: A-ASSOCIATE-AC
SCU->>SCP: C-ECHO-RQ
SCP-->>SCU: C-ECHO-RSP, status 0000
SCU->>SCP: A-RELEASE-RQ
SCP-->>SCU: A-RELEASE-RP
Every step in that diagram is something that can fail, so a green echo does establish something. It’s just narrower than it looks:
| A successful echo establishes | It still leaves open |
|---|---|
| A TCP connection to that host and port, from this machine, at this moment | Whether other senders, other network paths, or the same path under load can connect |
| The SCP accepted an association with this calling and called AE title | Whether it checks AE titles at all. Some SCPs accept any calling AE, and some answer to more than one called AE, so a green echo doesn’t prove you reached the application you meant to |
| The Verification presentation context was accepted | Whether any storage, query/retrieve, or worklist context will be |
| The Verification SCP processed a request | Whether the storage, database, indexing, or routing behind it is healthy |
| This AE can reach that one | Whether that one can open an association back to you |
It’s still a useful list. The first two rows cover most of what a network team or a vendor engineer means by “connectivity,” and echo answers them in under a second.
What it doesn’t touch#
Now look at what a real transfer does that echo does not.
Presentation context negotiation for the object you’re sending. Echo proposes Verification. Storage proposes CT Image Storage with specific transfer syntaxes. The SCP can accept the association and still reject the context you need, and it tells you so in the association accept, not with an error. Each proposed context comes back with a result: accepted, rejected by the user, rejected with no reason, abstract syntax not supported, or transfer syntaxes not supported (PS3.8, Section 9.3.3.2). A tool that only checks “did the association open” never sees it. An open association isn’t necessarily a usable one. This is the conformance statement failure made real: both sides speak DICOM, neither speaks the same SOP class, and echo says everything’s fine.
Transfer syntax agreement. Same mechanism, one level down. The standard requires the association requester to offer implicit VR little endian in at least one presentation context for every SOP class it proposes, and forbids the acceptor from rejecting all of them because it supports none of the transfer syntaxes (PS3.5, Section 10.1), so a conformant pair can fall back to uncompressed. The failures come from the exceptions and the non-conformers. 10.1 waives this for a sender that has the pixel data only in lossy-compressed form, and for one whose lossless-compressed or encapsulated uncompressed pixel data is too long to encode in the default transfer syntax (unless it offers a pixel-data-reference syntax); such a sender may propose only its compressed or encapsulated syntax. And some senders propose only their compressed syntax whatever the standard says. If the receiver doesn’t accept it, that context comes back rejected, normally as transfer syntaxes not supported, and echo is still green.
The data set. Echo sends none. Storage sends a full object, and now the receiving application has to deal with it. Missing required attributes, wrong VRs, values that contradict each other, pixel data that doesn’t match what the header says it holds: some SCPs check for these during the C-STORE and return an error or warning status; others accept the object and find the problem later in their own processing. A modality that produces a slightly malformed header, an anonymizer that stripped a Type 1 attribute, a transfer syntax header that lies about its contents: all of these can fail, at the C-STORE or later, on a connection that echoes perfectly.
Size and duration. Echo is a few hundred bytes and done in milliseconds. A large store runs for seconds or minutes and pushes megabytes to gigabytes through every network device between the two AEs. That exercises timeouts, long-lived TLS sessions and inspection appliances, buffering, and the receiver’s memory and disk. An echo over TLS proves the handshake succeeded and that each side accepted the other’s certificate as far as it is configured to check, but none of the rest. Maximum PDU length is negotiated in the association, and a correct sender splits a large object across as many PDUs as it needs, so a PDU setting is rarely the cause on its own. It’s worth reading in the association log, but in my experience the appliances and the timeouts are the likelier suspects.
Concurrency and load. One echo on one association. In one department I’ve supported, production meant twelve modalities on forty associations at shift change. The SCP’s association limit, its thread pool, and its disk behave differently under that load, and echo tests none of it.
Storage commitment and MPPS. A successful C-STORE response says the store succeeded. Storage commitment is a separate service that asks a separate question: will the archive take responsibility for keeping these objects? The sender issues an N-ACTION request, and the answer comes back as an N-EVENT-REPORT, which the SCP may send on a new association it opens back toward the sender (PS3.4, Annex J). On the modalities I’ve worked with, if deletion of local copies waits for commitment and the answer never arrives, the modality keeps them and eventually fills its disk. MPPS is a third service: the modality reports with N-CREATE and N-SET that a procedure started and then completed or was discontinued (PS3.4, Annex F). Each is its own SOP class on its own presentation context, and the echo proposed neither.
The return path. Echo is one-directional. Any workflow where the far side initiates (C-MOVE sub-operations, storage commitment events sent on a new association) needs the far side to be able to associate to you, which is a separate echo in the other direction that, in my experience, almost nobody runs. MPPS doesn’t have this problem, since the modality initiates every message in it.
What to run instead#
The sequence below is the one I settled on after building a few generations of test tools, first with the DCMTK command-line utilities and later in Python with pynetdicom. Each step assumes the one before it passed, and each one isolates a layer the previous step couldn’t see.
flowchart TD
A["1. Echo, plus the reverse
direction if the workflow needs it"] --> B["2. Associate with the real
SOP classes and transfer
syntaxes; inspect the accept"]
B --> C["3. Store one small,
known-good object"]
C --> D["4. Store one large object
(same SOP class, if you can)"]
D --> E["5. Store in parallel at
production concurrency,
then sustain it"]
E --> F["6. Query it back; confirm
what arrived is what left"]
F --> G["7. Commitment / MPPS
round trip, if in scope"]
classDef s fill:#1e293b,stroke:#64748b,color:#e2e8f0;
class A,B,C,D,E,F,G s;
1. Echo, and the reverse direction if the workflow needs it. Yes, still first. It’s cheap and it clears the network and AE-title layer so that everything after it is about DICOM. If anything in scope has the far side opening an association back to you (C-MOVE destinations, storage commitment replies), echo from there to your AE as well. A modality that only pushes images doesn’t need it.
2. Associate with the real contexts. Open an association proposing exactly the SOP classes and transfer syntaxes the production sender will propose, and read the accept. Every context comes back accepted or rejected with a reason. This single step catches the majority of go-live failures I’ve seen, and it takes a few seconds. DCMTK’s storescu with -d dumps the negotiated contexts and their results; pynetdicom gives you the accepted contexts as a list on the association object.
3. One small object. A known-good file, a few hundred kilobytes, in a transfer syntax you confirmed in step 2. If this fails, the problem is in the storage operation itself, not the network, the AE titles, or the negotiation. The C-STORE status, the association log, and the receiver’s own log are what tell a rejected data set apart from a full disk or an application error.
4. One large object. The biggest thing the modality produces: a tomosynthesis stack, an enhanced multi-frame MR, a long cine. Where you can, keep the SOP class and transfer syntax the same as in step 3, so size is the only thing that changed. This finds the path problems: timeouts, appliances that only inspect big flows, receivers that run out of something.
5. Concurrency. Repeat the stores across several simultaneous associations against the production port. My tester ramps through five, ten, fifteen and twenty concurrent senders; pick numbers that match your busiest hour. Watch for association refusals (the SCP’s limit), response times that grow faster than the load, and the occasional object that fails under load but not alone. Then hold the load long enough to expose a backlog or a leak that a one-minute burst won’t. This is the test that most closely predicts seven in the morning. Run it against a test system, or in a quiet window agreed with the people who depend on the production one.
6. Query it back. C-FIND for what you just stored, at the series or image level if the archive supports it, and compare the UIDs to what you sent; a study-level answer can’t tell you every instance arrived. A C-STORE success status confirms the store. It doesn’t promise the archive has indexed the object where a query will find it, and some archives hold or quarantine objects after the fact. When it matters that the pixels arrived intact, retrieve a few and compare. The C-FIND post covers what you can and can’t learn from that query.
7. The round trip. Storage commitment and MPPS, if the modality uses them. Commitment is where the reverse path from step 1 earns its keep, but the reverse echo only proves the network path and that the modality accepts that AE title pair. When the archive sends the result on a new association, which every commitment SCP must be able to do, it opens that association using the same AE title, proposing Storage Commitment with a role-selection item asking for the SCP role. The modality has to accept the context and return that role. Read that accept too. MPPS needs its own context accepted by whatever receives it.
Each step produces a yes or a no at one layer. A no at step 4 after a yes at step 3 narrows the search to what’s different about a large object before anyone opens a packet capture. That’s the whole value of the sequence: it localizes.
When echo is the right test#
To be fair to it. Echo is the right test when the question is “is the thing up,” and that’s a question worth asking often. A scheduled echo every few minutes against every AE you care about, with an alert on failure, is one of the cheapest pieces of monitoring an imaging department can run, and it catches the listener-is-down class of outage before the first tech calls. I’ve run it from a cron job and from a monitoring stack; either works.
Read it for what it tells you, though. A failed echo means something is wrong on that path. A green one doesn’t mean the archive is healthy: it can keep answering echo through a full disk, a stalled queue, or a broken index. Watch those separately.
It’s the wrong test when the question is “will it work,” and the two questions get confused because the answer to the first one arrives so fast and so green.
Sources#
- DICOM PS3.4, Annex A, Verification Service Class: https://dicom.nema.org/medical/dicom/current/output/chtml/part04/chapter_A.html.
- DICOM PS3.4, Annex J, Storage Commitment Service Class, and Annex F, Modality Performed Procedure Step: https://dicom.nema.org/medical/dicom/current/output/chtml/part04/chapter_J.html, https://dicom.nema.org/medical/dicom/current/output/chtml/part04/chapter_F.html.
- DICOM PS3.5, Section 10.1, for the default transfer syntax every implementation must support: https://dicom.nema.org/medical/dicom/current/output/chtml/part05/chapter_10.html.
- DICOM PS3.4, Section J.3.3.1, for the commitment result on a new association, the same AE title, and SCP role selection: https://dicom.nema.org/medical/dicom/current/output/chtml/part04/sect_J.3.3.html. DICOM PS3.7, Section D.3.3.4, SCP/SCU Role Selection Negotiation: https://dicom.nema.org/medical/dicom/current/output/chtml/part07/sect_D.3.3.4.html.
- DICOM PS3.7, Message Exchange, for C-ECHO, C-STORE, and status codes: https://dicom.nema.org/medical/dicom/current/output/chtml/part07/PS3.7.html.
- DICOM PS3.8, Network Communication Support, Section 9.3.3.2, for the presentation context results in the association accept: https://dicom.nema.org/medical/dicom/current/output/chtml/part08/sect_9.3.3.2.html.
- DCMTK: https://dicom.offis.de/dcmtk. pynetdicom: https://pydicom.github.io/pynetdicom/.
- The test sequence and the failure patterns are from my own tooling and engagements, and are stated as experience.
