diff --git a/doc/connection-backoff-interop-test-description.md b/doc/connection-backoff-interop-test-description.md
new file mode 100644
index 0000000000000000000000000000000000000000..0f00c86dca202a87619352b817abb7300bd985d1
--- /dev/null
+++ b/doc/connection-backoff-interop-test-description.md
@@ -0,0 +1,77 @@
+Connection Backoff Interop Test Descriptions
+===============================================
+
+This test is to verify the client is reconnecting the server with correct
+backoffs as specified in
+[the spec](http://github.com/grpc/grpc/blob/master/doc/connection-backoff.md).
+The test server has a port (control_port) running a rpc service for controlling
+the server and another port (retry_port) to close any incoming tcp connections.
+The test has the following flow:
+
+1. The server starts listening on control_port.
+2. The client calls Start rpc on server control_port.
+3. The server starts listening on retry_port.
+4. The client connects to server retry_port and retries with backoff for 540s,
+which translates to about 13 retries.
+5. The client calls Stop rpc on server control port.
+6. The client checks the response to see whether the server thinks the backoffs
+are conforming the spec or do its own check on the backoffs in the response.
+
+Client and server use
+[test.proto](https://github.com/grpc/grpc/blob/master/test/proto/test.proto).
+Each language should implement its own client. The C++ server is shared among
+languages.
+
+Client
+------
+
+Clients should accept these arguments:
+* --server_control_port=PORT
+    * The server port to connect to for rpc. For example, "8080"
+* --server_retry_port=PORT
+    * The server port to connect to for testing backoffs. For example, "8081"
+
+The client must connect to the control port without TLS. The client should
+either assert on the server returned backoff status or check the returned
+backoffs on its own.
+
+Procedure of client:
+
+1. Calls Start on server control port with a large deadline or no deadline,
+waits for its finish and checks it succeeded.
+2. Initiates a channel connection to server retry port, which should perform
+reconnections with proper backoffs. A convienent way to achieve this is to
+call Start with a deadline of 540s. The rpc should fail with deadline exceeded.
+3. Calls Stop on server control port and checks it succeeded.
+4. Checks the response to see whether the server thinks the backoffs passed the
+   test.
+5. Optionally, the client can do its own check on the returned backoffs.
+
+
+Server
+------
+
+A C++ server can be used for the test. Other languages do NOT need to implement
+a server. To minimize the network delay, the server binary should run on the
+same machine or on a nearby machine (in terms of network distance) with the
+client binary.
+
+A server implements the ReconnectService to its state. It also opens a
+tcp server on the retry_port, which just shuts down all incoming tcp
+connections to simulate connection failures. The server will keep a record of
+all the reconnection timestamps and return the connection backoffs in the
+response in milliseconds. The server also checks the backoffs to see whether
+they conform the spec and returns whether the client passes the test.
+
+If the server receives a Start call when another client is being tested, it
+finishes the call when the other client is done. If some other host connects
+to the server retry_port when a client is being tested, the server will log an
+error but likely would think the client fails the test.
+
+The server accepts these arguments:
+
+* --control_port=PORT
+    * The port to listen on for control rpcs. For example, "8080"
+* --retry_port=PORT
+    * The tcp server port. For example, "8081"
+