Decentralized counters for ETS ordered_set with write_concurrency

Previously, all ETS tables used centralized counter variables to keep
track of the number of items stored and the amount of memory
consumed. These counters can cause scalability problems (especially on
big NUMA systems). This commit adds an implementation of a
decentralized counter and modifies the implementation of ETS so that
ETS tables of type ordered_set with write_concurrency enabled use the
decentralized counter. [Experiments][1] indicate that this change
substantially improves the scalability of ETS ordered_set tables with
write_concurrency enabled in scenarios with frequent `ets:insert/2`
and `ets:delete/2` calls.

The new counter is implemented in the module erts_flxctr
(`erts_flxctr.h` and `erts_flxctr.c`). The module has the suffix
flxctr as it contains the implementation of a flexible counter (i.e.,
counter instances can be configured to be either centralized or
decentralized). Counters that are configured to be centralized are
implemented with a single counter variable which is modified with
atomic operations. Decentralized counters are spread over several
cache lines (how many can be configured with the parameter
`+dcg`). The scheduler threads are mapped to cache lines so that there
is no single point of contention when decentralized counters are
updated. The thread progress functionality of the Erlang VM is
utilized to implement support for linearizable snapshots of
decentralized counters. The snapshot functionality is used by the
`ets:info/1` and `ets:info/2` functions.

[1]: http://winsh.me/ets_catree_benchmark/flxctr_res.html

1 files changed, 406 insertions, 0 deletions

diff --git a/erts/emulator/beam/erl_flxctr.h b/erts/emulator/beam/erl_flxctr.h
new file mode 100644
index 0000000000..5cab02b9eb
--- /dev/null
+++ b/erts/emulator/beam/erl_flxctr.h
@@ -0,0 +1,406 @@
+/*
+ * %CopyrightBegin%
+ *
+ * Copyright Ericsson AB 2019. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ *
+ * %CopyrightEnd%
+ */
+
+/**
+ * @file erl_flxctr.h
+ *
+ * @brief This file contains the API of a flexible counter. The
+ * counter can be configured during its initialization to be
+ * centralized or decentralized. The centralized configuration makes
+ * it possible to read the counter value extremely efficiently, but
+ * updates of the counter value can easily cause contention. The
+ * decentralized configuration has the reverse trade-off (i.e.,
+ * updates are efficient and scalable but reading the counter value is
+ * slow and may cause contention).
+ *
+ * @author Kjell Winblad
+ */
+
+#ifndef ERL_FLXCTR_H__
+#define ERL_FLXCTR_H__
+
+#ifdef HAVE_CONFIG_H
+#  include "config.h"
+#endif
+#include "sys.h"
+#include "erl_vm.h"
+#include "global.h"
+#include "error.h"
+#include "bif.h"
+#include "big.h"
+#include "erl_binary.h"
+#include "bif.h"
+#include <stddef.h>
+
+/* Public Interface */
+
+#define ERTS_MAX_FLXCTR_GROUPS 256
+#define ERTS_FLXCTR_ATOMICS_PER_CACHE_LINE (ERTS_CACHE_LINE_SIZE / sizeof(erts_atomic_t))
+
+typedef struct {
+    int nr_of_counters;
+    int is_decentralized;
+    union {
+        erts_atomic_t counters_ptr;
+        erts_atomic_t counters[1];
+    } u;
+} ErtsFlxCtr;
+
+#define ERTS_FLXCTR_NR_OF_EXTRA_BYTES(NR_OF_COUNTERS)   \
+    ((NR_OF_COUNTERS-1) * sizeof(erts_atomic_t))
+
+/* Called by early_init */
+void erts_flxctr_setup(int decentralized_counter_groups);
+
+/**
+ * @brief Initializes an ErtsFlxCtr. The macro
+ * ERTS_FLXCTR_NR_OF_EXTRA_BYTES should be used to determine how much
+ * extra space that needs to be allocated directly after the
+ * ErtsFlxCtr when is_decentralized is set to zero.  Each ErtsFlxCtr
+ * instance may contain up to ERTS_FLXCTR_ATOMICS_PER_CACHE_LINE
+ * counters. These counters are numbered from zero to
+ * (ERTS_FLXCTR_ATOMICS_PER_CACHE_LINE-1). Most of the functions in
+ * this module take a parameter named counter_nr that controls which
+ * of the ERTS_FLXCTR_ATOMICS_PER_CACHE_LINE counters in the given
+ * ErtsFlxCtr that should be operated on.
+ *
+ * @param c The counter to initialize
+ * @param is_decentralized Non-zero value to make c decentralized
+ * @param nr_of_counters The number of counters included in c
+ *                       (max ERTS_FLXCTR_ATOMICS_PER_CACHE_LINE)
+ * @param alloc_type 
+ */
+void erts_flxctr_init(ErtsFlxCtr* c,
+                      int is_decentralized,
+                      Uint nr_of_counters,
+                      ErtsAlcType_t alloc_type);
+
+/**
+ * @brief Destroys an initialized counter.
+ *
+ * @param c The counter that should be destroyed
+ * @param alloc_type The allocation type (needs to be the same as the
+ *                   one passed to erts_flxctr_init when c was
+ *                   initialized)
+ */
+void erts_flxctr_destroy(ErtsFlxCtr* c, ErtsAlcType_t alloc_type);
+
+/**
+ * @brief Adds to_add to the counter with counter_nr in c
+ *
+ * @param c the ErtsFlxCtr to operate on
+ * @param counter_nr The number of the counter in c to modify
+ * @param to_add The amount that should be added to the specified counter
+ */
+ERTS_GLB_INLINE
+void erts_flxctr_add(ErtsFlxCtr* c,
+                     Uint counter_nr,
+                     int to_add);
+
+/**
+ * @brief Increases the specified counter by 1
+ *
+ * @param c The ErtsFlxCtr instance to operate on
+ * @param counter_nr The number of the counter within c to operate on
+ */
+ERTS_GLB_INLINE
+void erts_flxctr_inc(ErtsFlxCtr* c,
+                     Uint counter_nr);
+
+/**
+ * @brief Decreases the specified counter by 1
+ */
+ERTS_GLB_INLINE
+void erts_flxctr_dec(ErtsFlxCtr* c,
+                     Uint counter_nr);
+
+/**
+ * @brief This function tries to return the current value of the
+ * specified counter but may return an incorrect result if the counter
+ * is decentralized and other threads are accessing the counter
+ * concurrently.
+ *
+ * @param c The ErtsFlxCtr instance to operate on
+ * @param counter_nr The number of the counter within c to operate on
+ *
+ * @return A snapshot of the specifed counter if c is centralized or a
+ *         possibly incorrect estimate of the counter value if c is
+ *         decentralized
+ */
+Sint erts_flxctr_read_approx(ErtsFlxCtr* c,
+                             Uint counter_nr);
+
+/**
+ * @brief This function can only be used together with an ErtsFlxCtr
+ * that is configured to be centralized. The function increments the
+ * specified counter by 1 and returns the value of the counter after
+ * the increment.
+ */
+ERTS_GLB_INLINE
+Sint erts_flxctr_inc_read_centralized(ErtsFlxCtr* c,
+                                      Uint counter_nr);
+
+/**
+ * @brief This function can only be used together with a ErtsFlxCtr
+ * that is configured to be centralized. The function decrements the
+ * specified counter by 1 and returns the value of the counter after
+ * the operation.
+ */
+ERTS_GLB_INLINE
+Sint erts_flxctr_dec_read_centralized(ErtsFlxCtr* c,
+                                      Uint counter_nr);
+
+/**
+ * @brief This function can only be used together with an ErtsFlxCtr
+ * that is configured to be centralized. The function returns the
+ * current value of the specified counter.
+ */
+ERTS_GLB_INLINE
+Sint erts_flxctr_read_centralized(ErtsFlxCtr* c,
+                                  Uint counter_nr);
+
+
+typedef enum {
+    ERTS_FLXCTR_TRY_AGAIN_AFTER_TRAP,
+    ERTS_FLXCTR_DONE,
+    ERTS_FLXCTR_GET_RESULT_AFTER_TRAP
+} ErtsFlxctrSnapshotResultType;
+
+typedef struct {
+    ErtsFlxctrSnapshotResultType type;
+    Eterm trap_resume_state;
+    Sint result[ERTS_FLXCTR_ATOMICS_PER_CACHE_LINE];
+} ErtsFlxCtrSnapshotResult;
+
+/**
+ * @brief This function initiates an atomic snapshot of an ErtsFlxCtr
+ * to read out the values of one or more of the counters that are
+ * stored in the given ErtsFlxCtr. The caller needs to perform
+ * different actions after the return of this function depending on
+ * the value of the type field in the returned struct:
+ *
+ * - The caller needs to trap and try again after the trap if the
+ *   return value has the type ERTS_FLXCTR_TRY_AGAIN_AFTER_TRAP.
+ *
+ * - The caller can get the result directly from the result field of
+ *   the returned struct if the return value has the type
+ *   ERTS_FLXCTR_DONE. The value at index i in the result field
+ *   correspond to counter number i.
+ *
+ * - Finally, if the return value has the type
+ *   ERTS_FLXCTR_GET_RESULT_AFTER_TRAP, then the caller needs to save
+ *   the value of the field trap_resume_state from the returned struct
+ *   and trap. After the trap, the values of the counters can be
+ *   obtained by using the function
+ *   erts_flxctr_get_snapshot_result_after_trap. Note that the
+ *   function erts_flxctr_is_snapshot_result can be used to check if a
+ *   value is obtained from the trap_resume_state field in the
+ *   returned struct (this can be useful when the calling function
+ *   wakes up again after the trap).
+ *
+ * The snapshot operation that is initiated by this function should be
+ * considered to be ongoing from the issuing of this function until a
+ * struct with the type field set to ERTS_FLXCTR_DONE has been
+ * returned from the function or until the caller of this function has
+ * woken up after trapping.
+ *
+ * @param c The ErtsFlxCtr that the snapshot shall be taken from
+ * @param alloc_type The allocation type (needs to be the same as the
+ *                   type passed to erts_flxctr_init when c was
+ *                   initialized)
+ * @param p The Erlang process that is doing the call
+ *
+ * @return See the description above
+ *
+ */
+ErtsFlxCtrSnapshotResult
+erts_flxctr_snapshot(ErtsFlxCtr* c,
+                     ErtsAlcType_t alloc_type,
+                     Process* p);
+
+/**
+ * @brief Checks if the parameter term is a snapshot result (i.e.,
+ * something obtained from the trap_resume_state field of an
+ * ErtsFlxCtrSnapshotResult struct that has been returned from
+ * erts_flxctr_snapshot).
+ *
+ * @param term The term to check 
+ *
+ * @return A nonzero value iff the term is a snapshot result
+ */
+int erts_flxctr_is_snapshot_result(Eterm term);
+
+/**
+ * @brief Returns the result of a snapshot for a counter given a
+ * snapshot result returned by a call to erts_flxctr_snapshot (i.e.,
+ * the value stored in the trap_resume_state field of a struct
+ * returned by erts_flxctr_snapshot). The caller needs to trap between
+ * the return of erts_flxctr_snapshot and the call to this function.
+ */
+Sint erts_flxctr_get_snapshot_result_after_trap(Eterm trap_resume_state,
+                                                Uint counter_nr);
+
+/**
+ * @brief Resets the specified counter to 0. This function is unsafe
+ * to call while a snapshot operation may be active (initiated with
+ * the erts_flxctr_snapshot function).
+ */
+void erts_flxctr_reset(ErtsFlxCtr* c,
+                       Uint counter_nr);
+
+/**
+ * @brief Checks if a snapshot operation is active (snapshots are
+ * initiated with the erts_flxctr_snapshot function).
+ *
+ * @return nonzero value iff a snapshot was active at some point
+ * between the invocation and return of the function
+ */
+int erts_flxctr_is_snapshot_ongoing(ErtsFlxCtr* c);
+
+/**
+ * @brief This function checks if a snapshot operation is ongoing
+ * (snapshots are initiated with the erts_flxctr_snapshot function)
+ * and suspend the given process until thread progress has happened if
+ * it detected an ongoing snapshot operation. The caller needs to trap
+ * if a non-zero value is returned.
+ *
+ * @param c The ErtsFlxCtr to check
+ * @param p The calling process
+ *
+ * @return nonzero value if the given process has got suspended
+ */
+int erts_flxctr_suspend_until_thr_prg_if_snapshot_ongoing(ErtsFlxCtr* c, Process* p);
+
+/* End: Public Interface */
+
+/* Internal Declarations */
+
+#define ERTS_FLXCTR_GET_CTR_ARRAY_PTR(C)                                \
+    ((ErtsFlxCtrDecentralizedCtrArray*) erts_atomic_read_acqb(&(C)->u.counters_ptr))
+#define ERTS_FLXCTR_GET_CTR_PTR(C, SCHEDULER_ID, COUNTER_ID)            \
+    &(ERTS_FLXCTR_GET_CTR_ARRAY_PTR(C))->array[SCHEDULER_ID].counters[COUNTER_ID]
+
+
+typedef union {
+    erts_atomic_t counters[ERTS_FLXCTR_ATOMICS_PER_CACHE_LINE];
+    char pad[ERTS_CACHE_LINE_SIZE];
+} ErtsFlxCtrDecentralizedCtrArrayElem;
+
+typedef struct ErtsFlxCtrDecentralizedCtrArray {
+    void* block_start;
+    erts_atomic_t snapshot_status;
+    ErtsFlxCtrDecentralizedCtrArrayElem array[];
+} ErtsFlxCtrDecentralizedCtrArray;
+
+void erts_flxctr_set_slot(int group);
+
+ERTS_GLB_INLINE
+int erts_flxctr_get_slot_index(void);
+
+/* End: Internal Declarations */
+
+
+/* Implementation of inlined functions */
+
+#if ERTS_GLB_INLINE_INCL_FUNC_DEF
+
+ERTS_GLB_INLINE
+int erts_flxctr_get_slot_index(void)
+{
+    ErtsSchedulerData *esdp = erts_get_scheduler_data();
+    ASSERT(esdp && !ERTS_SCHEDULER_IS_DIRTY(esdp));
+    ASSERT(esdp->flxctr_slot_no > 0);
+    return esdp->flxctr_slot_no;
+}
+
+ERTS_GLB_INLINE
+void erts_flxctr_add(ErtsFlxCtr* c,
+                     Uint counter_nr,
+                     int to_add)
+{
+    ASSERT(counter_nr < c->nr_of_counters);
+    if (c->is_decentralized) {
+        erts_atomic_add_nob(ERTS_FLXCTR_GET_CTR_PTR(c,
+                                                    erts_flxctr_get_slot_index(),
+                                                    counter_nr),
+                            to_add);
+    } else {
+        erts_atomic_add_nob(&c->u.counters[counter_nr], to_add);
+    }
+}
+
+ERTS_GLB_INLINE
+void erts_flxctr_inc(ErtsFlxCtr* c,
+                     Uint counter_nr)
+{
+    ASSERT(counter_nr < c->nr_of_counters);
+    if (c->is_decentralized) {
+        erts_atomic_inc_nob(ERTS_FLXCTR_GET_CTR_PTR(c,
+                                                    erts_flxctr_get_slot_index(),
+                                                    counter_nr));
+    } else {
+        erts_atomic_inc_read_nob(&c->u.counters[counter_nr]);
+    }
+}
+
+ERTS_GLB_INLINE
+void erts_flxctr_dec(ErtsFlxCtr* c,
+                     Uint counter_nr)
+{
+    ASSERT(counter_nr < c->nr_of_counters);
+    if (c->is_decentralized) {
+        erts_atomic_dec_nob(ERTS_FLXCTR_GET_CTR_PTR(c,
+                                                    erts_flxctr_get_slot_index(),
+                                                    counter_nr));
+    } else {
+        erts_atomic_dec_nob(&c->u.counters[counter_nr]);
+    }
+}
+
+ERTS_GLB_INLINE
+Sint erts_flxctr_inc_read_centralized(ErtsFlxCtr* c,
+                                      Uint counter_nr)
+{
+    ASSERT(counter_nr < c->nr_of_counters);
+    ASSERT(!c->is_decentralized);
+    return erts_atomic_inc_read_nob(&c->u.counters[counter_nr]);
+}
+
+ERTS_GLB_INLINE
+Sint erts_flxctr_dec_read_centralized(ErtsFlxCtr* c,
+                                      Uint counter_nr)
+{
+    ASSERT(counter_nr < c->nr_of_counters);
+    ASSERT(!c->is_decentralized);
+    return erts_atomic_dec_read_nob(&c->u.counters[counter_nr]);
+}
+
+ERTS_GLB_INLINE
+Sint erts_flxctr_read_centralized(ErtsFlxCtr* c,
+                                  Uint counter_nr)
+{
+    ASSERT(counter_nr < c->nr_of_counters);
+    ASSERT(!c->is_decentralized);   
+    return erts_atomic_read_nob(&((erts_atomic_t*)(c->u.counters))[counter_nr]);
+}
+
+#endif /* #if ERTS_GLB_INLINE_INCL_FUNC_DEF */
+
+#endif /* ERL_FLXCTR_H__ */