diff --git a/backend/Cargo.toml b/backend/Cargo.toml
index 7061f61..21279c2 100644
--- a/backend/Cargo.toml
+++ b/backend/Cargo.toml
@@ -12,3 +12,4 @@ opt-level = 3
 [profile.release]
 lto = true
 panic = "abort"
+debug = true
\ No newline at end of file
diff --git a/backend/telemetry_core/tests/soak_tests.rs b/backend/telemetry_core/tests/soak_tests.rs
index 4201bff..6b91a08 100644
--- a/backend/telemetry_core/tests/soak_tests.rs
+++ b/backend/telemetry_core/tests/soak_tests.rs
@@ -16,6 +16,38 @@ sudo sysctl -w kern.ipc.maxsockbuf=16777216
 
 In general, if you run into issues, it may be better to run this on a linux
 box; MacOS seems to hit limits quicker in general.
+
+We can profile execution using cargo-flamegraph to learn more about how long
+we spend in each function.
+
+The main thing to do is install it:
+
+```sh
+cargo install flamegraph
+```
+
+Now, start processes running ourselves (so that we can use flamegraph on them):
+
+In one terminal (in the `backend` folder of this repo):
+
+```sh
+sudo cargo flamegraph -o telemetry_core.svg --bin telemetry_core
+```
+
+And in another:
+
+```sh
+sudo cargo flamegraph -o telemetry_shard.svg --bin telemetry_shard
+```
+
+And then, in a third terminal, we can run our general soak test against these:
+
+```sh
+SOAK_TEST_ARGS='--feeds 100 --nodes 100 --shards 4' \
+TELEMETRY_SUBMIT_HOSTS='127.0.0.1:8001' \
+TELEMETRY_FEED_HOST='127.0.0.1:8000' \
+cargo test -- soak_test --ignored --nocapture
+```
 */
 
 use futures::{ StreamExt };
diff --git a/backend/test_utils/src/server/server.rs b/backend/test_utils/src/server/server.rs
index 113f1fa..0ef9d4a 100644
--- a/backend/test_utils/src/server/server.rs
+++ b/backend/test_utils/src/server/server.rs
@@ -58,6 +58,11 @@ pub enum Server {
         core: CoreProcess,
     },
     ConnectToExistingMode {
+        /// The hosts that we can connect to to submit things.
+        submit_hosts: Vec<String>,
+        /// Which host do we use next (we'll cycle around them
+        /// as shards are "added").
+        next_submit_host_idx: usize,
         /// Shard processes that we can connect to.
         shards: DenseMap<ProcessId, ShardProcess>,
         /// Core process that we can connect to.
@@ -151,13 +156,27 @@ impl Server {
     /// Connect a new shard and return a process that you can interact with:
     pub async fn add_shard(&mut self) -> Result<ProcessId, Error> {
         match self {
-            // Always get back the same "shard" in virtual mode; it's just the core anyway.
+            // Always get back the same "virtual" shard; we're always just talking to the core anyway.
             Server::SingleProcessMode { virtual_shard, .. } => {
                 Ok(virtual_shard.id)
             },
-            // We're connecting to existing things; nothing sane to hand back.
-            Server::ConnectToExistingMode { .. } => {
-                Err(Error::CannotAddShard)
+            // We're connecting to an existing process. Find the next host we've been told about
+            // round-robin style and use that as our new virtual shard.
+            Server::ConnectToExistingMode { submit_hosts, next_submit_host_idx, shards, .. } => {
+                let host = match submit_hosts.get(*next_submit_host_idx % submit_hosts.len()) {
+                    Some(host) => host,
+                    None => return Err(Error::CannotAddShard)
+                };
+                *next_submit_host_idx += 1;
+
+                let pid = shards.add_with(|id| Process {
+                    id,
+                    host: format!("{}", host),
+                    handle: None,
+                    _channel_type: PhantomData,
+                });
+
+                Ok(pid)
             },
             // Start a new process and return that.
             Server::ShardAndCoreMode { shard_command, shards, core } => {
@@ -234,18 +253,10 @@ impl Server {
                 }
             },
             StartOpts::ConnectToExisting { feed_host, submit_hosts } => {
-                let mut shards = DenseMap::new();
-                for host in submit_hosts {
-                    shards.add_with(|id| Process {
-                        id,
-                        host,
-                        handle: None,
-                        _channel_type: PhantomData,
-                    });
-                }
-
                 Server::ConnectToExistingMode {
-                    shards,
+                    submit_hosts,
+                    next_submit_host_idx: 0,
+                    shards: DenseMap::new(),
                     core: Process {
                         id: ProcessId(0),
                         host: feed_host,
diff --git a/backend/test_utils/src/workspace/commands.rs b/backend/test_utils/src/workspace/commands.rs
index df7f97a..595b2b2 100644
--- a/backend/test_utils/src/workspace/commands.rs
+++ b/backend/test_utils/src/workspace/commands.rs
@@ -21,7 +21,7 @@ fn telemetry_command(bin: &'static str, release_mode: bool) -> Result<Command, s
     let mut workspace_dir = try_find_workspace_dir()?;
     workspace_dir.push("Cargo.toml");
 
-    let mut cmd = Command::new("cargo").arg("run");
+    let mut cmd = Command::new("cargo").arg("flamegraph");
 
     // Release mode?
     if release_mode {
diff --git a/backend/test_utils/src/workspace/start_server.rs b/backend/test_utils/src/workspace/start_server.rs
index 8229d4c..8bbf590 100644
--- a/backend/test_utils/src/workspace/start_server.rs
+++ b/backend/test_utils/src/workspace/start_server.rs
@@ -11,23 +11,40 @@ use crate::server::{self, Server, Command};
 /// - `TELEMETRY_SHARD_BIN` - path to telemetry_shard binary
 /// - `TELEMETRY_CORE_BIN` - path to telemetry_core binary
 ///
-/// Whatever is not provided will be substituted with a `cargo run` variant instead.
+/// (Whatever is not provided will be substituted with a `cargo run` variant instead)
+///
+/// Or alternately alternately, we can connect to a running instance by providing:
+///
+/// - `TELEMETRY_SUBMIT_HOSTS` - hosts (comma separated) to connect to for telemetry `/submit`s.
+/// - `TELEMETRY_FEED_HOST` - host to connect to for feeds (eg 127.0.0.1:3000)
+///
 pub async fn start_server(release_mode: bool) -> Server {
-
+    // Start to a single process:
     if let Ok(bin) = std::env::var("TELEMETRY_BIN") {
         return Server::start(server::StartOpts::SingleProcess {
             command: Command::new(bin)
         }).await.unwrap();
     }
 
+    // Connect to a running instance:
+    if let Ok(feed_host) = std::env::var("TELEMETRY_FEED_HOST") {
+        let feed_host = feed_host.trim().into();
+        let submit_hosts: Vec<_> = std::env::var("TELEMETRY_SUBMIT_HOSTS")
+            .map(|var| var.split(",").map(|var| var.trim().into()).collect())
+            .unwrap_or(Vec::new());
+        return Server::start(server::StartOpts::ConnectToExisting {
+            feed_host,
+            submit_hosts,
+        }).await.unwrap();
+    }
+
+    // Start a shard and core process:
     let shard_command = std::env::var("TELEMETRY_SHARD_BIN")
         .map(|val| Command::new(val))
         .unwrap_or_else(|_| commands::cargo_run_telemetry_shard(release_mode).expect("must be in rust workspace to run shard command"));
-
     let core_command = std::env::var("TELEMETRY_CORE_BIN")
         .map(|val| Command::new(val))
         .unwrap_or_else(|_| commands::cargo_run_telemetry_core(release_mode).expect("must be in rust workspace to run core command"));
-
     Server::start(server::StartOpts::ShardAndCore {
         shard_command,
         core_command