Age | Commit message (Collapse) | Author |
|
This commit only affects the implementation of ETS `ordered_set`
tables with the `write_concurrency` option enabled. Such tables are
implemented with a data structure that is called the contention
adapting search tree (CA tree). This commit introduces the following
changes:
* This commit causes a join to be triggered in one randomly selected
base node in about one of 1000 read unlock calls for base node
locks. No such joins happened before this commit. Before this
commit, operations that only acquired looks in read-mode never
triggered any contention adaptation. Therefore, the CA tree could
get stuck in a sub-optimal state in certain scenarios. This could
happen, for example, when a CA tree is first populated with parallel
inserts (which will cause splits of base nodes) and then only
read-only operations are applied to the data structure. Benchmark
results from the
`ets_SUITE:lookup_catree_par_vs_seq_init_benchmark/0` benchmark
function (which is included in this commit) shows that this change
can improve the throughput of the CA tree in the scenario described
above.
* Read-only operations will now also increase values of statistics
counters when they detect that they need to wait for other
operations. Only write operation changed statistics counters before
this commit. This improves the statistics that the adaptation
heuristics is based on.
* Additionally, this commit adds an upper and lower limit to the
contention statistics variables in the base nodes. Such limits did
not exist before this commit. This should, for example, make the CA
tree more responsive to contention after long periods of low
contention.
|
|
This commit fixes an ETS test case that tests the decentralized memory
counter in tables of type ordered_set with the write_concurrency
option turned on. The test case assumed that the memory consumption of
the table would only grow monotonically when terms are
inserted. However, this was not the case when the emulator was
compiled in debug mode as random splits and joins of CA tree nodes
could happen. This commit fixes the test case by disabling random
splits and joins in the tested table.
|
|
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
|
|
We no longer lock more than one base node at a time.
We do however trylock a second base node at join.
|
|
Once an iteration key has been found, never fall back to first/last key in
next/prev tree as trees may split or join under our feet. I.e we must always
use previous key when searching for the next key.
|
|
to easier generate a routing tree for test
without having to spend cpu to provoke actual repeated lock conflicts.
|
|
{RouteNodes, BaseNodes, MaxRouteTreeDepth}
|
|
|
|
to not abuse DbTerm.
|
|
Lock order is reverse tree depth, from leafs toward root.
This solution may eventually fail if running too long
as route nodes do not increase their 'lc_order' in a join operation
when they move up in the tree.
But who runs a VM with lock-checker for such a long time?
|
|
Lock order is key term order, so each base node needs its own key
if lock check is enabled.
|
|
Update table memory stats before scheduling free
to not be dependent on deallocation order with main table struct.
|
|
u as in union
|
|
The current ETS ordered_set implementation can quickly become a
scalability bottleneck on multicore machines when an application updates
an ordered_set table from concurrent processes [1][2]. The current
implementation is based on an AVL tree protected from concurrent writes
by a single readers-writer lock. Furthermore, the current implementation
has an optimization, called the stack optimization [3], that can improve
the performance when only a single process accesses a table but can
cause bad scalability even in read-only scenarios. It is possible to
pass the option {write_concurrency, true} to ets:new/2 when creating an
ETS table of type ordered_set but this option has no effect for tables
of type ordered_set without this commit. The new ETS ordered_set
implementation, added by this commit, is only activated when one passes
the options ordered_set and {write_concurrency, true} to the ets:new/2
function. Thus, the previous ordered_set implementation (from here on
called the default implementation) can still be used in applications
that do not benefit from the new implementation. The benchmark results
on the following web page show that the new implementation is many times
faster than the old implementation in some scenarios and that the old
implementation is still better than the new implementation in some
scenarios.
http://winsh.me/ets_catree_benchmark/ets_ca_tree_benchmark_results.html
The new implementation is expected to scale better than the default
implementation when concurrent processes use the following ETS
operations to operate on a table:
delete/2, delete_object/2, first/1, insert/2 (single object),
insert_new/2 (single object), lookup/2, lookup_element/2, member/2,
next/2, take/2 and update_element/3 (single object).
Currently, the new implementation does not have scalable support for the
other operations (e.g., select/2). However, when these operations are
used infrequently, the new implantation may still scale better than the
default implementation as the benchmark results at the URL above shows.
Description of the New Implementation
----------------------------------
The new implementation is based on a data structure which is called the
contention adapting search tree (CA tree for short). The following
publication contains a detailed description of the CA tree:
A Contention Adapting Approach to Concurrent Ordered Sets
Journal of Parallel and Distributed Computing, 2018
Kjell Winblad and Konstantinos Sagonas
https://doi.org/10.1016/j.jpdc.2017.11.007
http://www.it.uu.se/research/group/languages/software/ca_tree/catree_proofs.pdf
A discussion of how the CA tree can be used as an ETS back-end can be
found in another publication [1]. The CA tree is a data structure that
dynamically changes its synchronization granularity based on detected
contention. Internally, the CA tree uses instances of a sequential data
structure to store items. The CA tree implementation contained in this
commit uses the same AVL tree implementation as is used for the default
ordered set implementation. This AVL tree implementation is reused so
that much of the existing code to implement the ETS operations can be
reused.
Tests
-----
The ETS tests in `lib/stdlib/test/ets_SUITE.erl` have been extended to
also test the new ordered_set implementation. The function
ets_SUITE:throughput_benchmark/0 has also been added to this file. This
function can be used to measure and compare the performance of the
different ETS table types and options. This function writes benchmark
data to standard output that can be visualized by the HTML page
`lib/stdlib/test/ets_SUITE_data/visualize_throughput.html`.
[1]
More Scalable Ordered Set for ETS Using Adaptation.
In Thirteenth ACM SIGPLAN workshop on Erlang (2014).
Kjell Winblad and Konstantinos Sagonas.
https://doi.org/10.1145/2633448.2633455
http://www.it.uu.se/research/group/languages/software/ca_tree/erlang_paper.pdf
[2]
On the Scalability of the Erlang Term Storage
In Twelfth ACM SIGPLAN workshop on Erlang (2013)
Kjell Winblad, David Klaftenegger and Konstantinos Sagonas
https://doi.org/10.1145/2505305.2505308
http://winsh.me/papers/erlang_workshop_2013.pdf
[3]
The stack optimization works by keeping one preallocated stack instance
in every ordered_set table. This stack is updated so that it contains
the search path in some read operations (e.g., ets:next/2). This makes
it possible for a subsequent ets:next/2 to avoid traversing some nodes
in some cases. Unfortunately, the preallocated stack needs to be flagged
so that it is not updated concurrently by several threads which cause
bad scalability.
|