Almost every outbound message is serialized to buffers of 16k pagesize.
We were serializing these messages off the IO loop (and retaining the concrete message
instance as well) and would then enqueue it on the IO loop to be dealt with as soon as the
channel is ready.
1. This would cause buffers to be held onto for longer than necessary, causing less reuse on average.
2. If a channel was slow for some reason, not only would concrete message instances queue up for it, but also 16k of buffers would be reserved for each message until it would be written+flushed physically.
With this change, the serialization happens on the event loop which effectively limits the number of buffers that `N` IO-threads will ever use so long as messages are small and channels writable.
Also, this change dereferences the reference to the concrete outbound message as soon as it has been serialized to save some more on GC.
This reduces the GC time for a default PMC run by about 50% in experiments (3 nodes, 2G heap each, loopback ... obvious caveat is that GC isn't that heavy in the first place with recent changes but still a measurable gain).
I also expect it to be helpful for master node stability by causing less of a spike if master is e.g. hit by a large number of requests that are processed batched (e.g. shard snapshot status updates) and responded to in a short time frame all at once.
Obviously, the downside to this change is that it introduces more latency on the IO loop for the serialization. But since we read all of these messages on the IO loop as well I don't see it as much of a qualitative change really and the more predictable buffer use seems much more valuable relatively.
We don't need to hold on to the request body past the beginning of sending
the response. There is no need to keep a reference to it until after the response
has been sent fully and we can eagerly release it here.
Note, this can be optimized further to release the contents even earlier but for now
this is an easy increment to saving some memory on the IO pool.
Elasticsearch requires that a HttpRequest abstraction be implemented
by http modules before server processing. This abstraction controls when
underlying resources are released. This commit moves this abstraction to
be created immediately after content aggregation. This change will
enable follow-up work including moving Cors logic into the server
package and tracking bytes as they are aggregated from the network
level.
In most cases we are seeing a `PooledHeapByteBuf` here now. No need to
redundantly create an new `ByteBuffer` and single element array for it
here when we can just directly unwrap its internal `byte[]`.
We never do any file IO or other blocking work on the transport threads
so no tangible benefit can be derived from using more threads than CPUs
for IO.
There are however significant downsides to using more threads than necessary
with Netty in particular. Since we use the default setting for
`io.netty.allocator.useCacheForAllThreads` which is `true` we end up
using up to `16MB` of thread local buffer cache for each transport thread.
Meaning we potentially waste CPUs * 16MB of heap for unnecessary IO threads in addition to obvious inefficiencies of artificially adding extra context switches.
Currently Elasticsearch creates independent event loop groups for each
transport (http and internal) transport type. This is unnecessary and
can lead to contention when different threads access shared resources
(ex: allocators). This commit moves to a model where, by default, the
event loops are shared between the transports. The previous behavior can
be attained by specifically setting the http worker count.
Currently there is a clear mechanism to stub sending a request through
the transport. However, this is limited to testing exceptions on the
sender side. This commit reworks our transport related testing
infrastructure to allow stubbing request handling on the receiving side.
This is a first cut at giving NodeInfo the ability to carry a flexible
list of heterogeneous info responses. The trick is to be able to
serialize and deserialize an arbitrary list of blocks of information. It
is convenient to be able to deserialize into usable Java objects so that
we can aggregate nodes stats for the cluster stats endpoint.
In order to provide a little bit of clarity about which objects can and
can't be used as info blocks, I've introduced a new interface called
"ReportingService."
I have removed the hard-coded getters (e.g., getOs()) in favor of a
flexible method that can return heterogeneous kinds of info blocks
(e.g., getInfo(OsInfo.class)). Taking a class as an argument removes the
need to cast in the client code.
The use of available processors, the terminology, and the settings
around it have evolved over time. This commit cleans up some places in
the codes and in the docs to adjust to the current terminology.
This commit moves the action name validation and circuit breaking into
the InboundAggregator. This work is valuable because it lays the
groundwork for incrementally circuit breaking as data is received.
This PR includes the follow behavioral change:
Handshakes contribute to circuit breaking, but cannot be broken. They
currently do not contribute nor are they broken.
* Refactor nodes stats request builders to match requests (#54363)
* Remove hard-coded setters from NodesInfoRequestBuilder
* Remove hard-coded setters from NodesStatsRequest
* Use static imports to reduce clutter
* Remove uses of old info APIs
This is a simple naming change PR, to fix the fact that "metadata" is a
single English word, and for too long we have not followed general
naming conventions for it. We are also not consistent about it, for
example, METADATA instead of META_DATA if we were trying to be
consistent with MetaData (although METADATA is correct when considered
in the context of "metadata"). This was a simple find and replace across
the code base, only taking a few minutes to fix this naming issue
forever.
Currently all of our transport protocol decoding and aggregation occurs
in the individual transport modules. This means that each implementation
(test, netty, nio) must implement this logic. Additionally, it means
that the entire message has been read from the network before the server
package receives it.
This commit creates a pipeline in server which can be passed arbitrary
bytes to handle. Internally, the pipeline will decode, decompress, and
aggregate the messages. Additionally, this allows us to run many
megabytes of bytes through the pipeline in tests to ensure that the
logic works.
This work will enable future work:
Circuit breaking or backoff logic based on message type and byte
in the content aggregator.
Sharing bytes with the application layer using the ref counted
releasable network bytes.
Improved network monitoring based specifically on channels.
Finally, this fixes the bug where we do not circuit break on the correct
message size when compression is enabled.
Elasticsearch has a number of different BytesReference implementations.
These implementations can all implement the interface in different ways
with subtly different behavior and performance characteristics. On the
other-hand, the JVM only represents bytes as an array or a direct byte
buffer. This commit deletes the specialized Netty implementations and
moves to using a generic ByteBuffer reference type. This will allow us
to focus on standardizing performance and behave around a smaller number
of implementations that can be used by all components in Elasticsearch.
Backport of #52542.
This commit is part of issue #40366 to remove disabled Xlint warnings
from gradle files. In particular, it removes the Xlint exclusions from
the following files:
- benchmarks/build.gradle
- client/client-benchmark-noop-api-plugin/build.gradle
- x-pack/qa/rolling-upgrade/build.gradle
- x-pack/qa/third-party/active-directory/build.gradle
- modules/transport-netty4/build.gradle
For the first three files no code adjustments were needed. For
x-pack/qa/third-party/active-directory move the suppression at the code
level. For transport-netty4 replace the variable arguments with
ArrayLists and remove any redundant casts.
Now that the FIPS 140 security provider is simply a test dependency
we don't need the thirdPartyAudit exceptions, but plugin-cli and
transport-netty4 do need jarHell disabled as they use the non fips
BouncyCastle security provider as a test dependency too.
rest-api-spec/test/10_basic.yml would check that transport_types is
`netty4` but we run FIPS 140 tests with default distribution and
transport_types is `security4`
This change changes the way to run our test suites in
JVMs configured in FIPS 140 approved mode. It does so by:
- Configuring any given runtime Java in FIPS mode with the bundled
policy and security properties files, setting the system
properties java.security.properties and java.security.policy
with the == operator that overrides the default JVM properties
and policy.
- When runtime java is 11 and higher, using BouncyCastle FIPS
Cryptographic provider and BCJSSE in FIPS mode. These are
used as testRuntime dependencies for unit
tests and internal clusters, and copied (relevant jars)
explicitly to the lib directory for testclusters used in REST tests
- When runtime java is 8, using BouncyCastle FIPS
Cryptographic provider and SunJSSE in FIPS mode.
Running the tests in FIPS 140 approved mode doesn't require an
additional configuration either in CI workers or locally and is
controlled by specifying -Dtests.fips.enabled=true
* Stop Allocating Buffers in CopyBytesSocketChannel (#49825)
The way things currently work, we read up to 1M from the channel
and then potentially force all of it into the `ByteBuf` passed
by Netty. Since that `ByteBuf` tends to by default be `64k` in size,
large reads will force the buffer to grow, completely circumventing
the logic of `allocHandle`.
This seems like it could break
`io.netty.channel.RecvByteBufAllocator.Handle#continueReading`
since that method for the fixed-size allocator does check
whether the last read was equal to the attempted read size.
So if we set `64k` because that's what the buffer size is,
then wirte `1M` to the buffer we will stop reading on the IO loop,
even though the channel may still have bytes that we can read right away.
More imporatantly though, this can lead to running OOM quite easily
under IO pressure as we are forcing the heap buffers passed to the read
to `reallocate`.
Closes#49699
* Copying the request is not necessary here. We can simply release it once the response has been generated and a lot of `Unpooled` allocations that way
* Relates #32228
* I think the issue that preventet that PR that PR from being merged was solved by #39634 that moved the bulk index marker search to ByteBuf bulk access so the composite buffer shouldn't require many additional bounds checks (I'd argue the bounds checks we add, we save when copying the composite buffer)
* I couldn't neccessarily reproduce much of a speedup from this change, but I could reproduce a very measureable reduction in GC time with e.g. Rally's PMC (4g heap node and bulk requests of size 5k saw a reduction in young GC time by ~10% for me)
The default merge cumulator used in netty transport leads to additional
GC pressure and memory copying when a message that exceeds the chunk
size is handled. This is especially a problem on G1 GC, since we get
many "humongous" allocations and that can in theory cause real memory
circuit breaker to break unnecessarily.
Backport of #48849. Update `.editorconfig` to make the Java settings the
default for all files, and then apply a 2-space indent to all `*.gradle`
files. Then reformat all the files.
This commit introduces a consistent, and type-safe manner for handling
global build parameters through out our build logic. Primarily this
replaces the existing usages of extra properties with static accessors.
It also introduces and explicit API for initialization and mutation of
any such parameters, as well as better error handling for uninitialized
or eager access of parameter values.
Closes#42042
The 1MB IO-buffer size per transport thread is causing trouble in
some tests, albeit at a low rate. Reducing the number of transport
threads was not enough to fully fix this situation.
Allowing to configure the size of the buffer and reducing it by
more than an order of magnitude should fix these tests.
Closes#46803
BytesReference is currently an abstract class which is extended by
various implementations. This makes it very difficult to use the
delegation pattern. The implication of this is that our releasable
BytesReference is a PagedBytesReference type and cannot be used as a
generic releasable bytes reference that delegates to any reference type.
This commit makes BytesReference an interface and introduces an
AbstractBytesReference for common functionality.
This commit removes the randomization used by every execute call in the
high level rest tests. Previously every execute call, which can be many
calls per single test, would rely on a random boolean to determine if
they should use the sync or async methods provided to the execute
method. This commit runs the tests twice, using two different clusters,
both of them providing the value one time via a sysprop. This ensures
that the whole suite of tests is run using the sync and async code
paths.
Closes#39667
This commit removes a problematic assertion that the netty default
allocator is not used. This assertion is problematic because any other
test can cause this task to fail by touching the default allocator. We
assert that we are using heap buffers in the channel.
This commit removes the option to change the netty system properties to
reenable the direct buffer pooling. It also removes the need for us to
disable the buffer pooling in the system properties file. Instead, we
programmatically craete an allocator that is used by our networking
layer.
This commit does introduce an Elasticsearch property which allows the
user to fallback on the netty default allocator. If they choose this
option, they can configure the default allocator how they wish using the
standard netty properties.
Currently in production instances of Elasticsearch we set a couple of
system properties by default. We currently do not apply all of these
system properties in tests. This commit applies these properties in the
tests.
Some netty behavior is controlled by system properties. While we want to
test with the defaults for Elasticsearch for most tests, within netty we
want to ensure these netty settings exhibit correct behavior. This
commit adds variants of test and integTest tasks for netty which set the
unpooled and direct buffer pooled allocators.
relates #45881
Currently we use a custom CopyBytesSocketChannel for interfacing with
netty. We have integration tests that use this channel, however we never
verify the read and write behavior in the face of potential partial
writes. This commit adds a test for this behavior.
This commit namespaces the existing processors setting under the "node"
namespace. In doing so, we deprecate the existing processors setting in
favor of node.processors.
Currently we take the array of nio buffers from the netty channel
outbound buffer and copy their bytes to a direct buffer. In the process
we mutate the nio buffer positions. It seems like netty will continue to
reuse these buffers. This means than any data that is not flushed in a
call is lost. This commit fixes this by incrementing the positions after
the flush has completed. This is similar to the behavior that
SocketChannel would have provided and netty relied upon.
Fixes#45444.
Elasticsearch does not grant Netty reflection access to get Unsafe. The
only mechanism that currently exists to free direct buffers in a timely
manner is to use Unsafe. This leads to the occasional scenario, under
heavy network load, that direct byte buffers can slowly build up without
being freed.
This commit disables Netty direct buffer pooling and moves to a strategy
of using a single thread-local direct buffer for interfacing with sockets.
This will reduce the memory usage from networking. Elasticsearch
currently derives very little value from direct buffer usage (TLS,
compression, Lucene, Elasticsearch handling, etc all use heap bytes). So
this seems like the correct trade-off until that changes.
Uses JDK 11's per-socket configuration of TCP keepalive (supported on Linux and Mac), see
https://bugs.openjdk.java.net/browse/JDK-8194298, and exposes these as transport settings.
By default, these options are disabled for now (i.e. fall-back to OS behavior), but we would like
to explore whether we can enable them by default, in particular to force keepalive configurations
that are better tuned for running ES.
Currently in the transport-nio work we connect and bind channels on the
a thread before the channel is registered with a selector. Additionally,
it is at this point that we set all the socket options. This commit
moves these operations onto the event-loop after the channel has been
registered with a selector. It attempts to set the socket options for a
non-server channel at registration time. If that fails, it will attempt
to set the options after the channel is connected. This should fix
#41071.
* Stop Passing Around REST Request in Multiple Spots
* Motivated by #44564
* We are currently passing the REST request object around to a large number of places. This works fine since we simply copy the full request content before we handle the rest itself which is needlessly hard on GC and heap.
* This PR removes a number of spots where the request is passed around needlessly. There are many more spots to optimize in follow-ups to this, but this one would already enable bypassing the request copying for some error paths in a follow up.
Armin Braun
Tim Brooks
William Brafford
Jason Tedor
Jake Landis
Rory Hunter
Maria Ralli
Ioannis Kakavas
Yannick Welsch
Henning Andersen
Rory Hunter
Mark Vieira
Yogesh Gaikwad
Michael Basnight
Ryan Ernst