WIP: work on fast multi-acquisition

Meaning, work towards multiple acquisitions after another with minimal delay inbetween.
LiberTEM · Nov 15, 2023 · 989705c · 989705c
1 parent ff40f94
commit 989705c
Show file tree

Hide file tree

Showing 27 changed files with 971 additions and 482 deletions.
diff --git a/Cargo.lock b/Cargo.lock
diff --git a/k2o/Cargo.toml b/k2o/Cargo.toml
@@ -31,6 +31,7 @@ log = "0.4.18"
 dbus = "0.9.7"
 libc = "0.2.147"
 env_logger = "0.10.0"
+partialdebug = "0.2.0"
 
 [features]
 hdf5 = ["dep:hdf5"]

diff --git a/k2o/NOTES.md b/k2o/NOTES.md
@@ -19,6 +19,38 @@
     - => no change in performance, still spending 88% of our time in the page fault handler
 - assembly of full frames is required to get a good access pattern!
 
+# Notes on reworking the life cycles
+
+To make sure we can quickly start acquisitions after another,
+we need to make sure of some boundary conditions:
+
+- The SHM area needs to be kept alive over multiple acquisitions, as well as
+  the "recycling" system for blocks, such that we don't have malloc/pagefault
+  caused jitter
+- The sockets need to be kept open, otherwise the multicast traffic may only be
+  received after an initial delay (IIRC)
+- That also means that switching between more fundamental settings may take longer
+    - changing network settings
+    - changing camera mode (IS/Summit)
+    - SHM socket path
+    - enabling / disabling the frame iterator
+- What changes must be fast:
+    - file writer destination (filename) -> we may need to disable pre-allocation of files!
+    - number of frames per acquisition
+    - probably: camera sync mode (immediately vs. wait for sync flag)
+
+## Action items
+- [ ] Refactor...
+    - [x] The `AcquisitionRuntime` should be the object that lives over multiple acquisitions
+    - [x] On the Python side, the `Cam` already needs to start the `AcquisitionRuntime`
+    - [ ] For error handling, we should have the option to completely tear down
+      and re-create the `AcquisitionRuntime`, perform this step automatically such that
+      we don't have "restart the {script,notebook,app,server}" situations as often as currently
+    - [x] `WriterBuilder` must be set per acquisition, not when starting the runtime
+- [x] Figure out where the acquisition ID should be generated
+    - probably in the `AcquisitionRuntime`, as that is the long-living object
+      which manages the background thread(s) etc.
+
 # TODO
 - [ ] fix busy waiting (one of the `try_recv` loops, I guess)
 - [ ] fix subframe shtuff
diff --git a/k2o/benches/assemble_benchmarks.rs b/k2o/benches/assemble_benchmarks.rs
@@ -14,8 +14,8 @@ fn criterion_benchmark(c: &mut Criterion) {
 
     const TOTAL_INPUT_SIZE: usize = 1 * PACKET_SIZE;
 
-    let block = K2ISBlock::empty(42);
-    let mut frame: K2ISFrame = K2ISFrame::empty::<K2ISBlock>(42);
+    let block = K2ISBlock::empty(42, 0);
+    let mut frame: K2ISFrame = K2ISFrame::empty::<K2ISBlock>(42, 0);
 
     let mut assign_block = c.benchmark_group("assign_block* functions");
     assign_block.measurement_time(Duration::new(10, 0));

diff --git a/k2o/benches/decode_benchmarks.rs b/k2o/benches/decode_benchmarks.rs
@@ -152,7 +152,7 @@ fn criterion_benchmark(c: &mut Criterion) {
                 let in_chunks = input.chunks_exact(PACKET_SIZE);
 
                 for chunk in in_chunks {
-                    K2ISBlock::from_bytes(black_box(chunk), 0);
+                    K2ISBlock::from_bytes(black_box(chunk), 0, 0);
                 }
             })
         },