Merge pull request #383 from liulanzheng/main

update doc for lsmt lookup

Author: BigVan

README.md

@@ -12,7 +12,7 @@ Overlaybd is based on [PhotonLibOS](https://github.com/alibaba/PhotonLibOS), whi
 
 Overlaybd has 2 core component:
 * **Overlaybd**
-  is a block-device based image format, provideing a merged view of a sequence of block-based layers as a virtual block device.
+  is a block-device based image format, provideing a merged view of a sequence of block-based layers as a virtual block device. The LBA lookup algorithm employs a linearized B+ tree and AVX-512 to optimize performance, significantly accelerating search speed up to 10X. [Lookup Performance](https://github.com/containerd/overlaybd/blob/main/docs/lsmt_lookup.md)
 
 * **Zfile**
   is a compression file format which support seekalbe online decompression.

docs/README.md

@@ -28,6 +28,8 @@ Now this service contains an implementation of overlaybd based on [TCMU](https:/
 
 This service is based on [PhotonLibOS](https://github.com/alibaba/PhotonLibOS), which is a high-efficiency LibOS framework.
 
+The LBA lookup algorithm employs a linearized B+ tree and AVX-512 to optimize performance, significantly accelerating search speed up to 10X. [Lookup Performance](https://github.com/containerd/overlaybd/blob/main/docs/lsmt_lookup.md)
+
 ## Accelerated container image
 
 [GitHub](https://github.com/containerd/accelerated-container-image)

docs/lsmt_lookup.md

@@ -0,0 +1,15 @@
+# Lookup Algorithm in LSMT
+
+## Description
+
+LBA lookup in LSMT can be abstracted as a segment search problem, searching within a sorted set of non-overlapping intervals. Previously, we used binary search via std::lower_bound. Now, we've adopted a linearized B+ tree combined with AVX-512, which better exploits CPU cache efficiency and delivers over a 10X speedup in lookup performance. Even in environments without AVX-512 support, using a loop optimized with bitmask still yields significant performance gains.
+
+
+## Performance
+
+| segment count | b+tree + avx512 | b+tree + loop + bitmask | lower bound |
+|---------------|-----------------|---------------|-------------|
+| 1k   | 220 M/s | 42.2 M/s | 18.3 M/s |
+| 10k  | 160 M/s | 30.7 M/s | 12.8 M/s |
+| 100k | 108 M/s | 21.8 M/s | 8.6 M/s  |
+| 1M   | 57.4 M/s | 15.2 M/s | 5.6 M/s  |

src/overlaybd/lsmt/index.cpp

@@ -74,7 +74,12 @@ static constexpr uint32_t LEVEL_START_ID[MAX_LEVEL] =  {0,  8,  80,  728,  6560,
 
 struct DefaultInnerSearch {
     static uint32_t inner_search(const uint64_t *base, uint64_t x) {
-        return std::upper_bound(base, base + 8, x) - base;
+        uint8_t mask = 0;
+    #pragma GCC unroll 20
+        for (uint32_t i = 0; i < ORDER; i++) {
+            mask |= ( (base[i] <= x) << i );
+        }
+        return __builtin_popcount(mask);
     }
 };