feat: Add allocation logging and visualization (#64)

* Basic Logging in binary format * write correct Segment size * minor * Fix None unwrap * First visualization of allocations * Storage saving binary logging * faster bitmap getting * minor * visualize bitmap by storage layer * Speed up plotting by precalculating things * New get with output parameter, initial timestep is 1 * Experimented with unpacked storage * Enforce max timestep * Remove simple plot * visual improvements * First export to mp4 * minor * Color disks seperately * Validation helpers * minor * Calculate and Plot fragmentation * Indicate Timestep, handle empty storage * Type hints, formatting * parallel fragmentation building * Code reorganization * parallel video exporting * Rewrite of visualization script with better performance, memory usage and organization * backend selection parameter for matplotlib * Handle missing input file correctly * Plot correct global fragmentation * runtime plot adjustments * remove dependency of slider for exporting * Plot failed allocations * CLI improvements * plot free blocks * use packed bits for faster plotting * plot allocation sizes * minor changes for visual clarity * minor * minor * unaligned allocation tries * lines to seperate segments * allocation_log feature flag * more complete allocation_log feature flag * runtime allocation_log path * information on how to visualize allocations * moved scripts to project root * removed allocation tries, instead count cycles Different allocators have different or even no notion of a try. So instead remove tries completely and introduce an allocator agnostic metric to measure how long an allocation takes. * proportion of cycles spent in allocator
parcio · Jan 13, 2025 · 268a2ac · 268a2ac
1 parent 03effd6
commit 268a2ac
Show file tree

Hide file tree

Showing 8 changed files with 1,518 additions and 15 deletions.
diff --git a/betree/Cargo.toml b/betree/Cargo.toml
@@ -83,4 +83,6 @@ figment_config = ["figment"]
 latency_metrics = []
 experimental-api = []
 nvm = ["pmdk"]
+# Log the allocations and deallocations done for later analysis
+allocation_log = []
 
diff --git a/betree/src/allocator.rs b/betree/src/allocator.rs
@@ -1,9 +1,10 @@
 //! This module provides `SegmentAllocator` and `SegmentId` for bitmap
 //! allocation of 1GiB segments.
 
-use crate::{cow_bytes::CowBytes, storage_pool::DiskOffset, vdev::Block};
+use crate::{cow_bytes::CowBytes, storage_pool::DiskOffset, vdev::Block, Error};
 use bitvec::prelude::*;
 use byteorder::{BigEndian, ByteOrder};
+use std::io::Write;
 
 /// 256KiB, so that `vdev::BLOCK_SIZE * SEGMENT_SIZE == 1GiB`
 pub const SEGMENT_SIZE: usize = 1 << SEGMENT_SIZE_LOG_2;
@@ -55,7 +56,7 @@ impl SegmentAllocator {
             }
         };
         self.mark(offset, size, Action::Allocate);
-        Some(offset)
+        return Some(offset);
     }
 
     /// Allocates a block of the given `size` at `offset`.

diff --git a/betree/src/data_management/dmu.rs b/betree/src/data_management/dmu.rs
@@ -6,7 +6,7 @@ use super::{
     CopyOnWriteEvent, Dml, HasStoragePreference, Object, ObjectReference,
 };
 use crate::{
-    allocator::{Action, SegmentAllocator, SegmentId},
+    allocator::{Action, SegmentAllocator, SegmentId, SEGMENT_SIZE},
     buffer::Buf,
     cache::{Cache, ChangeKeyError, RemoveError},
     checksum::{Builder, Checksum, State},
@@ -17,16 +17,21 @@ use crate::{
     size::{Size, SizeMut, StaticSize},
     storage_pool::{DiskOffset, StoragePoolLayer, NUM_STORAGE_CLASSES},
     tree::{Node, PivotKey},
-    vdev::{Block, BLOCK_SIZE},
+    vdev::{Block, File, BLOCK_SIZE},
     StoragePreference,
 };
+use byteorder::{LittleEndian, WriteBytesExt};
 use crossbeam_channel::Sender;
 use futures::{executor::block_on, future::ok, prelude::*};
 use parking_lot::{Mutex, RwLock, RwLockReadGuard, RwLockWriteGuard};
 use std::{
+    arch::x86_64::{__rdtscp, _rdtsc},
     collections::HashMap,
+    fs::OpenOptions,
+    io::{BufWriter, Write},
     mem::replace,
     ops::DerefMut,
+    path::PathBuf,
     pin::Pin,
     sync::{
         atomic::{AtomicU64, Ordering},
@@ -60,6 +65,8 @@ where
     next_modified_node_id: AtomicU64,
     next_disk_id: AtomicU64,
     report_tx: Option<Sender<DmlMsg>>,
+    #[cfg(feature = "allocation_log")]
+    allocation_log_file: Mutex<BufWriter<std::fs::File>>,
 }
 
 impl<E, SPL> Dmu<E, SPL>
@@ -76,6 +83,7 @@ where
         alloc_strategy: [[Option<u8>; NUM_STORAGE_CLASSES]; NUM_STORAGE_CLASSES],
         cache: E,
         handler: Handler<ObjRef<ObjectPointer<SPL::Checksum>>>,
+        #[cfg(feature = "allocation_log")] allocation_log_file_path: PathBuf,
     ) -> Self {
         let allocation_data = (0..pool.storage_class_count())
             .map(|class| {
@@ -87,6 +95,16 @@ where
             .collect::<Vec<_>>()
             .into_boxed_slice();
 
+        #[cfg(feature = "allocation_log")]
+        let allocation_log_file = Mutex::new(BufWriter::new(
+            OpenOptions::new()
+                .create(true)
+                .write(true)
+                .truncate(true)
+                .open(allocation_log_file_path)
+                .expect("Failed to create allocation log file"),
+        ));
+
         Dmu {
             // default_compression_state: default_compression.new_compression().expect("Can't create compression state"),
             default_compression,
@@ -103,6 +121,8 @@ where
             next_modified_node_id: AtomicU64::new(1),
             next_disk_id: AtomicU64::new(0),
             report_tx: None,
+            #[cfg(feature = "allocation_log")]
+            allocation_log_file,
         }
     }
 
@@ -120,6 +140,36 @@ where
     pub fn pool(&self) -> &SPL {
         &self.pool
     }
+
+    /// Writes the global header for the allocation logging.
+    pub fn write_global_header(&self) -> Result<(), Error> {
+        #[cfg(feature = "allocation_log")]
+        {
+            let mut file = self.allocation_log_file.lock();
+
+            // Number of storage classes
+            file.write_u8(self.pool.storage_class_count())?;
+
+            // Disks per class
+            for class in 0..self.pool.storage_class_count() {
+                let disk_count = self.pool.disk_count(class);
+                file.write_u16::<LittleEndian>(disk_count)?;
+            }
+
+            // Segments per disk
+            for class in 0..self.pool.storage_class_count() {
+                for disk in 0..self.pool.disk_count(class) {
+                    let segment_count = self.pool.size_in_blocks(class, disk);
+                    file.write_u64::<LittleEndian>(segment_count.as_u64())?;
+                }
+            }
+
+            // Blocks per segment (constant)
+            file.write_u64::<LittleEndian>(SEGMENT_SIZE.try_into().unwrap())?;
+        }
+
+        Ok(())
+    }
 }
 
 impl<E, SPL> Dmu<E, SPL>
@@ -201,6 +251,15 @@ where
             obj_ptr.offset().disk_id(),
             obj_ptr.size(),
         );
+        #[cfg(feature = "allocation_log")]
+        {
+            let mut file = self.allocation_log_file.lock();
+            let _ = file.write_u8(Action::Deallocate.as_bool() as u8);
+            let _ = file.write_u64::<LittleEndian>(obj_ptr.offset.as_u64());
+            let _ = file.write_u32::<LittleEndian>(obj_ptr.size.as_u32());
+            let _ = file.write_u64::<LittleEndian>(0);
+            let _ = file.write_u64::<LittleEndian>(0);
+        }
         if let (CopyOnWriteEvent::Removed, Some(tx), CopyOnWriteReason::Remove) = (
             self.handler.copy_on_write(
                 obj_ptr.offset(),
@@ -484,6 +543,16 @@ where
 
         let strategy = self.alloc_strategy[storage_preference as usize];
 
+        // NOTE: Could we mark classes, disks and/or segments as full to prevent looping over them?
+        // We would then also need to handle this, when deallocating things.
+        // Would full mean completely full or just not having enough contiguous memory of some
+        // size?
+        // Or save the largest contiguous memory region as a value and compare against that. For
+        // that the allocator needs to support that and we have to 'bubble' the largest value up.
+        #[cfg(feature = "allocation_log")]
+        let mut start_cycles_global = get_cycles();
+        #[cfg(feature = "allocation_log")]
+        let mut total_cycles_local: u64 = 0;
         'class: for &class in strategy.iter().flatten() {
             let disks_in_class = self.pool.disk_count(class);
             if disks_in_class == 0 {
@@ -536,14 +605,40 @@ where
 
                 let first_seen_segment_id = *segment_id;
                 loop {
-                    if let Some(segment_offset) = self
-                        .handler
-                        .get_allocation_bitmap(*segment_id, self)?
-                        .access()
-                        .allocate(size.as_u32())
+                    // Has to be split because else the temporary value is dropped while borrowing
+                    let bitmap = self.handler.get_allocation_bitmap(*segment_id, self)?;
+                    let mut allocator = bitmap.access();
+
+                    #[cfg(not(feature = "allocation_log"))]
+                    {
+                        let allocation = allocator.allocate(size.as_u32());
+                        if let Some(segment_offset) = allocation {
+                            let disk_offset = segment_id.disk_offset(segment_offset);
+                            break disk_offset;
+                        }
+                    }
+                    #[cfg(feature = "allocation_log")]
                     {
-                        break segment_id.disk_offset(segment_offset);
+                        let start_cycles_allocation = get_cycles();
+                        let allocation = allocator.allocate(size.as_u32());
+                        let end_cycles_allocation = get_cycles();
+                        total_cycles_local += end_cycles_allocation - start_cycles_allocation;
+
+                        if let Some(segment_offset) = allocation {
+                            let disk_offset = segment_id.disk_offset(segment_offset);
+                            let total_cycles_global = end_cycles_allocation - start_cycles_global;
+
+                            let mut file = self.allocation_log_file.lock();
+                            file.write_u8(Action::Allocate.as_bool() as u8)?;
+                            file.write_u64::<LittleEndian>(disk_offset.as_u64())?;
+                            file.write_u32::<LittleEndian>(size.as_u32())?;
+                            file.write_u64::<LittleEndian>(total_cycles_local)?;
+                            file.write_u64::<LittleEndian>(total_cycles_global)?;
+
+                            break disk_offset;
+                        }
                     }
+
                     let next_segment_id = segment_id.next(disk_size);
                     trace!(
                         "Next allocator segment: {:?} -> {:?} ({:?})",
@@ -1031,3 +1126,11 @@ where
         self.report_tx = Some(tx);
     }
 }
+
+fn get_cycles() -> u64 {
+    unsafe {
+        //let mut aux = 0;
+        //__rdtscp(aux)
+        _rdtsc()
+    }
+}
diff --git a/betree/src/database/mod.rs b/betree/src/database/mod.rs
@@ -31,7 +31,7 @@ use serde::{de::DeserializeOwned, Deserialize, Serialize};
 use std::{
     collections::HashMap,
     iter::FromIterator,
-    path::Path,
+    path::{Path, PathBuf},
     sync::{
         atomic::{AtomicU64, Ordering},
         Arc,
@@ -147,6 +147,9 @@ pub struct DatabaseConfiguration {
 
     /// If and how to log database metrics
     pub metrics: Option<MetricsConfiguration>,
+
+    /// Where to log the allocations
+    pub allocation_log_file_path: PathBuf,
 }
 
 impl Default for DatabaseConfiguration {
@@ -162,6 +165,7 @@ impl Default for DatabaseConfiguration {
             sync_interval_ms: Some(DEFAULT_SYNC_INTERVAL_MS),
             metrics: None,
             migration_policy: None,
+            allocation_log_file_path: PathBuf::from("allocation_log.bin"),
         }
     }
 }
@@ -237,6 +241,8 @@ impl DatabaseConfiguration {
             strategy,
             ClockCache::new(self.cache_size),
             handler,
+            #[cfg(feature = "allocation_log")]
+            self.allocation_log_file_path.clone(),
         )
     }
 
@@ -432,6 +438,9 @@ impl Database {
             dmu.set_report(tx.clone());
         }
 
+        #[cfg(feature = "allocation_log")]
+        dmu.write_global_header()?;
+
         let (tree, root_ptr) = builder.select_root_tree(Arc::new(dmu))?;
 
         *tree.dmu().handler().current_generation.lock_write() = root_ptr.generation().next();

diff --git a/betree/src/tree/imp/mod.rs b/betree/src/tree/imp/mod.rs
@@ -393,14 +393,11 @@ where
                     self.msg_action().apply(key, &msg, &mut tmp);
                 }
 
-                // This may never be false.
-                let data = tmp.unwrap();
-
                 drop(node);
                 if self.evict {
                     self.dml.evict()?;
                 }
-                Ok(Some((info, data)))
+                Ok(tmp.map(|data| (info, data)))
             }
         }
     }

diff --git a/scripts/README.md b/scripts/README.md
@@ -0,0 +1,37 @@
+# Allocation Log Visualization
+
+This script visualizes the allocation and deallocation of blocks within the key-value database. It helps to understand how storage space is being used and identify potential optimization opportunities.
+
+The allocation log visualization script is tested with Python 3.12.7 and the packages listed in `requirements.txt`.
+
+The main dependencies are matplotlib, tqdm and sortedcontainers.
+
+## Setup
+
+Run the following to create a working environment for the script:
+
+```bash
+python3 -m venv .venv
+source .venv/bin/activate
+python3 -m pip install -r scripts/requirements.txt
+```
+
+## Generating the Allocation Log
+
+To generate the `allocation_log.bin` file, you need to enable the allocation_log feature flag when compiling the `betree` crate. For instance by running
+```bash
+cargo build --features allocation_log
+```
+or by enabling it in the `Cargo.toml`.
+
+The path where the log is saved can be set with the runtime configuration parameter `allocation_log_file_path`. The default is `$PWD/allocation_log.bin`
+
+## Using the Allocation Log
+
+Once a log file has been obtained simply run the following to visualize the (de-)allocations recorded.
+```bash
+./scripts/visualize_allocation_log allocation_log.bin
+```
+
+To get help and see the options available run the script with the `-h` flag.
+
diff --git a/scripts/requirements.txt b/scripts/requirements.txt
@@ -0,0 +1,13 @@
+contourpy==1.3.1
+cycler==0.12.1
+fonttools==4.55.3
+kiwisolver==1.4.7
+matplotlib==3.9.3
+numpy==2.2.0
+packaging==24.2
+pillow==11.0.0
+pyparsing==3.2.0
+python-dateutil==2.9.0.post0
+six==1.17.0
+sortedcontainers==2.4.0
+tqdm==4.67.1