os/bluestore: implement allocator fragmentation histogram

Signed-off-by: Igor Fedotov <igor.fedotov@croit.io>
(cherry picked from commit 625e3461fbed904e85dfeea7ab3e8f7b74c4f384)

diff --git a/src/os/bluestore/Allocator.cc b/src/os/bluestore/Allocator.cc
index 5725e720d6f56bf24c670d3039dc6770e81f7025..5c5b8db70eccdd502c38effe6454b219e2031ff0 100644 (file)
--- a/src/os/bluestore/Allocator.cc
+++ b/src/os/bluestore/Allocator.cc
@@ -14,6 +14,7 @@
 #include "common/debug.h"
 #include "common/admin_socket.h"
 #define dout_subsys ceph_subsys_bluestore
+using TOPNSPC::common::cmd_getval;
 
 using std::string;
 using std::to_string;
@@ -52,6 +53,13 @@ public:
           this,
           "give allocator fragmentation (0-no fragmentation, 1-absolute fragmentation)");
         ceph_assert(r == 0);
+        r = admin_socket->register_command(
+         ("bluestore allocator fragmentation histogram " + name +
+           " name=alloc_unit,type=CephInt,req=false" +
+           " name=num_buckets,type=CephInt,req=false").c_str(),
+         this,
+         "build allocator free regions state histogram");
+        ceph_assert(r == 0);
       }
     }
   }
@@ -100,6 +108,39 @@ public:
       f->open_object_section("fragmentation");
       f->dump_float("fragmentation_rating", alloc->get_fragmentation());
       f->close_section();
+    } else if (command == "bluestore allocator fragmentation histogram " + name) {
+      int64_t alloc_unit = 4096;
+      cmd_getval(cmdmap, "alloc_unit", alloc_unit);
+      if (alloc_unit == 0  ||
+          p2align(alloc_unit, alloc->get_block_size()) != alloc_unit) {
+        ss << "Invalid allocation unit: '" << alloc_unit
+           << ", to be aligned with: '" << alloc->get_block_size()
+           << std::endl;
+        return -EINVAL;
+      }
+      int64_t num_buckets = 8;
+      cmd_getval(cmdmap, "num_buckets", num_buckets);
+      if (num_buckets < 2) {
+        ss << "Invalid amount of buckets (min=2): '" << num_buckets
+           << std::endl;
+        return -EINVAL;
+      }
+
+      Allocator::FreeStateHistogram hist;
+      hist.resize(num_buckets);
+      alloc->build_free_state_histogram(alloc_unit, hist);
+      f->open_array_section("extent_counts");
+      for(int i = 0; i < num_buckets; i++) {
+        f->open_object_section("c");
+        f->dump_unsigned("max_len",
+          hist[i].get_max(i, num_buckets)
+        );
+        f->dump_unsigned("total", hist[i].total);
+        f->dump_unsigned("aligned", hist[i].aligned);
+        f->dump_unsigned("units", hist[i].alloc_units);
+        f->close_section();
+      }
+      f->close_section();
     } else {
       ss << "Invalid command" << std::endl;
       r = -ENOSYS;
@@ -234,3 +275,51 @@ double Allocator::get_fragmentation_score()
   double terrible = (sum / block_size) * get_score(block_size);
   return (ideal - score_sum) / (ideal - terrible);
 }
+
+void Allocator::build_free_state_histogram(
+  size_t alloc_unit, Allocator::FreeStateHistogram& hist)
+{
+  auto num_buckets = hist.size();
+  ceph_assert(num_buckets);
+
+  auto base = free_state_hist_bucket::base;
+  auto base_bits = free_state_hist_bucket::base_bits;
+  auto mux = free_state_hist_bucket::mux;
+  // maximum chunk size we track,
+  // provided by the bucket before the last one
+  size_t max =
+    free_state_hist_bucket::get_max(num_buckets - 2, num_buckets);
+
+  auto iterated_allocation = [&](size_t off, size_t len) {
+    size_t idx;
+    if (len <= base) {
+      idx = 0;
+    } else if (len > max) {
+      idx = num_buckets - 1;
+    } else {
+      size_t most_bit = cbits(uint64_t(len-1)) - 1;
+      idx = 1 + ((most_bit - base_bits) / mux);
+    }
+    ceph_assert(idx < num_buckets);
+    ++hist[idx].total;
+
+    // now calculate the bucket for the chunk after alignment,
+    // resulting chunks shorter than alloc_unit are discarded
+    auto delta = p2roundup(off, alloc_unit) - off;
+    if (len >= delta + alloc_unit) {
+      len -= delta;
+      if (len <= base) {
+        idx = 0;
+      } else if (len > max) {
+        idx = num_buckets - 1;
+      } else {
+        size_t most_bit = cbits(uint64_t(len-1)) - 1;
+        idx = 1 + ((most_bit - base_bits) / mux);
+      }
+      ++hist[idx].aligned;
+      hist[idx].alloc_units += len / alloc_unit;
+    }
+  };
+
+  foreach(iterated_allocation);
+}

diff --git a/src/os/bluestore/Allocator.h b/src/os/bluestore/Allocator.h
index e378007c3c0036a90619a22be04e805ce54cbe2e..f136c98b2926ad0e763f55fae7bcbfe75549b399 100644 (file)
--- a/src/os/bluestore/Allocator.h
+++ b/src/os/bluestore/Allocator.h
@@ -88,6 +88,37 @@ public:
     return block_size;
   }
 
+  // The following code build Allocator's free extents histogram.
+  // Which is a set of N buckets to track extents layout.
+  // Extent matches a bucket depending on its length using the following
+  // length spans:
+  // [0..4K] (4K..16K] (16K..64K] .. (4M..16M] (16M..]
+  // Each bucket tracks:
+  // - total amount of extents of specific lengths
+  // - amount of extents aligned with allocation boundary
+  // - amount of allocation units in aligned extents
+  //
+  struct free_state_hist_bucket {
+    static const size_t base_bits = 12;
+    static const size_t base = 1ull << base_bits;
+    static const size_t mux = 2;
+
+    size_t total = 0;
+    size_t aligned = 0;
+    size_t alloc_units = 0;
+
+    // returns upper bound of the bucket
+    static size_t get_max(size_t bucket, size_t num_buckets) {
+      return
+        bucket < num_buckets - 1 ?
+          base << (mux * bucket) :
+          std::numeric_limits<uint64_t>::max();
+    };
+  };
+
+  typedef std::vector<free_state_hist_bucket> FreeStateHistogram;
+  void build_free_state_histogram(size_t alloc_unit, FreeStateHistogram& hist);
+
 private:
   class SocketHook;
   SocketHook* asok_hook = nullptr;