From: Igor Fedotov <ifedotov@suse.com>
Date: Wed, 26 Feb 2020 21:01:49 +0000 (+0300)
Subject: os/bluestore: log allocation stats on a daily basis.
X-Git-Tag: v15.1.1~119^2
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=8a33d9c1dd5574da10ced345c5aab8cbe53d6961;p=ceph.git

os/bluestore: log allocation stats on a daily basis.
The primary goal is to be able to track allocator fragmentation trend.
Which might help in troubleshooting performance changes caused by aging.

Tracks history for up to last 31 days by keeping a set of 5 historic
probes selected from power of two sequence. I.e.
Dey -1
Day -2
Day -4
Day -8
Day -16

Signed-off-by: Igor Fedotov <ifedotov@suse.com>
---

diff --git a/src/common/legacy_config_opts.h b/src/common/legacy_config_opts.h
index 7aff6bedcade..fbe387230bf2 100644
--- a/src/common/legacy_config_opts.h
+++ b/src/common/legacy_config_opts.h
@@ -1000,6 +1000,7 @@ OPTION(bluestore_cache_size_hdd, OPT_U64)
 OPTION(bluestore_cache_size_ssd, OPT_U64)
 OPTION(bluestore_cache_meta_ratio, OPT_DOUBLE)
 OPTION(bluestore_cache_kv_ratio, OPT_DOUBLE)
+OPTION(bluestore_alloc_stats_dump_interval, OPT_DOUBLE)
 OPTION(bluestore_kvbackend, OPT_STR)
 OPTION(bluestore_allocator, OPT_STR)     // stupid | bitmap
 OPTION(bluestore_freelist_blocks_per_key, OPT_INT)
diff --git a/src/common/options.cc b/src/common/options.cc
index 4442e97ba822..0883e04cd600 100644
--- a/src/common/options.cc
+++ b/src/common/options.cc
@@ -4355,6 +4355,10 @@ std::vector<Option> get_global_options() {
     .add_see_also("bluestore_cache_autotune")
     .set_description("The number of seconds to wait between rebalances when cache autotune is enabled."),
 
+    Option("bluestore_alloc_stats_dump_interval", Option::TYPE_FLOAT, Option::LEVEL_DEV)
+      .set_default(3600 * 24)
+      .set_description("The period (in second) for logging allocation statistics."),
+
     Option("bluestore_kvbackend", Option::TYPE_STR, Option::LEVEL_DEV)
     .set_default("rocksdb")
     .set_flag(Option::FLAG_CREATE)
diff --git a/src/os/bluestore/BlueStore.cc b/src/os/bluestore/BlueStore.cc
index 6d2088d8a4fa..c6c5157ce0c4 100644
--- a/src/os/bluestore/BlueStore.cc
+++ b/src/os/bluestore/BlueStore.cc
@@ -3826,6 +3826,7 @@ void *BlueStore::MempoolThread::entry()
   utime_t next_balance = ceph_clock_now();
   utime_t next_resize = ceph_clock_now();
   utime_t next_deferred_force_submit = ceph_clock_now();
+  utime_t alloc_stats_dump_clock = ceph_clock_now();
 
   bool interval_stats_trim = false;
   while (!stop) {
@@ -3841,6 +3842,14 @@ void *BlueStore::MempoolThread::entry()
     double resize_interval = store->osd_memory_cache_resize_interval;
     double max_defer_interval = store->max_defer_interval;
 
+    double alloc_stats_dump_interval =
+      store->cct->_conf->bluestore_alloc_stats_dump_interval;
+
+    if (alloc_stats_dump_interval > 0 &&
+        alloc_stats_dump_clock + alloc_stats_dump_interval < ceph_clock_now()) {
+      store->_record_allocation_stats();
+      alloc_stats_dump_clock = ceph_clock_now();
+    }
     if (autotune_interval > 0 && next_balance < ceph_clock_now()) {
       _adjust_cache_settings();
 
@@ -3881,6 +3890,8 @@ void *BlueStore::MempoolThread::entry()
       store->cct->_conf->bluestore_cache_trim_interval);
     cond.wait_for(l, wait);
   }
+  // do final dump
+  store->_record_allocation_stats();
   stop = false;
   return NULL;
 }
@@ -5840,6 +5851,8 @@ int BlueStore::allocate_bluefs_freespace(
     uint64_t gift;
     uint64_t allocated = 0;
     int64_t alloc_len;
+    auto need = size;
+    auto extent_count0 = extents->size();
     do {
       // hard cap to fit into 32 bits
       gift = std::min<uint64_t>(size, 1ull << 30);
@@ -5869,6 +5882,8 @@ int BlueStore::allocate_bluefs_freespace(
 	return -ENOSPC;
       }
     } while (size && alloc_len > 0);
+    _collect_allocation_stats(need, alloc_size, extents->size() - extent_count0);
+
     for (auto& e : *extents) {
       dout(5) << __func__ << " gifting " << e << " to bluefs" << dendl;
       bluefs_extents.insert(e.offset, e.length);
@@ -13438,6 +13453,7 @@ int BlueStore::_do_alloc_write(
     }
     return -ENOSPC;
   }
+  _collect_allocation_stats(need, min_alloc_size, prealloc.size());
 
   dout(20) << __func__ << " prealloc " << prealloc << dendl;
   auto prealloc_pos = prealloc.begin();
@@ -15429,6 +15445,57 @@ void BlueStore::_log_alerts(osd_alert_list_t& alerts)
   }
 }
 
+void BlueStore::_collect_allocation_stats(uint64_t need, uint32_t alloc_size,
+                                          size_t extents)
+{
+  alloc_stats_count++;
+  alloc_stats_fragments += extents;
+  alloc_stats_size += need;
+}
+
+void BlueStore::_record_allocation_stats()
+{
+  // don't care about data consistency,
+  // fields can be partially modified while making the tuple
+  auto t0 = std::make_tuple(
+    alloc_stats_count.exchange(0),
+    alloc_stats_fragments.exchange(0),
+    alloc_stats_size.exchange(0));
+
+  dout(0) << " allocation stats probe "
+    << probe_count << ":"
+    << " cnt: " << std::get<0>(t0)
+    << " frags: " << std::get<1>(t0)
+    << " size: " << std::get<2>(t0)
+    << dendl;
+
+
+  //
+  // Keep the history for probes from the power-of-two sequence:
+  // -1, -2, -4, -8, -16
+  // 
+  size_t base = 1;
+  for (auto& t : alloc_stats_history) {
+    dout(0) << " probe -"
+      << base + (probe_count % base) << ": "
+      << std::get<0>(t)
+      << ",  " << std::get<1>(t)
+      << ", " << std::get<2>(t)
+      << dendl;
+    base <<= 1;
+  }
+  dout(0) << "------------" << dendl;
+
+  auto prev = probe_count++;
+  auto mask = (1 << alloc_stats_history.size()) - 1;
+  probe_count &= mask;
+
+  for (size_t i = cbits(prev ^ probe_count) - 1; i > 0 ; --i) {
+    alloc_stats_history[i] = alloc_stats_history[i - 1];
+  }
+  alloc_stats_history[0].swap(t0);
+}
+
 // ===========================================
 // BlueStoreRepairer
 
diff --git a/src/os/bluestore/BlueStore.h b/src/os/bluestore/BlueStore.h
index 544034a04bcc..1430f64fbea9 100644
--- a/src/os/bluestore/BlueStore.h
+++ b/src/os/bluestore/BlueStore.h
@@ -3206,7 +3206,18 @@ private:
 			CollectionRef& d,
 			unsigned bits);
 
+  void _collect_allocation_stats(uint64_t need, uint32_t alloc_size,
+                                 size_t extents);
+  void _record_allocation_stats();
 private:
+  uint64_t probe_count = 0;
+  std::atomic<uint64_t> alloc_stats_count = {0};
+  std::atomic<uint64_t> alloc_stats_fragments = { 0 };
+  std::atomic<uint64_t> alloc_stats_size = { 0 };
+  // 
+  std::array<std::tuple<uint64_t, uint64_t, uint64_t>, 5> alloc_stats_history =
+  { std::make_tuple(0ul, 0ul, 0ul) };
+
   std::atomic<uint64_t> out_of_sync_fm = {0};
   // --------------------------------------------------------
   // BlueFSDeviceExpander implementation