This package contains the canonical implementation of the libp2p Network Resource Manager interface.
The implementation is based on the concept of Resource Management Scopes, whereby resource usage is constrained by a DAG of scopes, accounting for multiple levels of resource constraints.
The Resource Manager doesn't prioritize resource requests at all, it simply checks if the resource being requested is currently below the defined limits and returns an error if the limit is reached. It has no notion of honest vs bad peers.
The Resource Manager does have a special notion of allowlisted multiaddrs that have their own limits if the normal system limits are reached.
The Resource Manager is intended to be used with go-libp2p. go-libp2p sets up a resource manager with the default autoscaled limits if none is provided, but if you want to configure things or if you want to enable metrics you'll use the resource manager like so:
// Start with the default scaling limits.
scalingLimits := rcmgr.DefaultLimits
// Add limits around included libp2p protocols
libp2p.SetDefaultServiceLimits(&scalingLimits)
// Turn the scaling limits into a concrete set of limits using `.AutoScale`. This
// scales the limits proportional to your system memory.
scaledDefaultLimits := scalingLimits.AutoScale()
// Tweak certain settings
cfg := rcmgr.PartialLimitConfig{
System: rcmgr.ResourceLimits{
// Allow unlimited outbound streams
StreamsOutbound: rcmgr.Unlimited,
},
// Everything else is default. The exact values will come from `scaledDefaultLimits` above.
}
// Create our limits by using our cfg and replacing the default values with values from `scaledDefaultLimits`
limits := cfg.Build(scaledDefaultLimits)
// The resource manager expects a limiter, se we create one from our limits.
limiter := rcmgr.NewFixedLimiter(limits)
// Metrics are enabled by default. If you want to disable metrics, use the
// WithMetricsDisabled option
// Initialize the resource manager
rm, err := rcmgr.NewResourceManager(limiter, rcmgr.WithMetricsDisabled())
if err != nil {
panic(err)
}
// Create a libp2p host
host, err := libp2p.New(libp2p.ResourceManager(rm))The easiest way to save the defined limits is to serialize the PartialLimitConfig
type as JSON.
noisyNeighbor, _ := peer.Decode("QmVvtzcZgCkMnSFf2dnrBPXrWuNFWNM9J3MpZQCvWPuVZf")
cfg := rcmgr.PartialLimitConfig{
System: &rcmgr.ResourceLimits{
// Allow unlimited outbound streams
StreamsOutbound: rcmgr.Unlimited,
},
Peer: map[peer.ID]rcmgr.ResourceLimits{
noisyNeighbor: {
// No inbound connections from this peer
ConnsInbound: rcmgr.BlockAllLimit,
// But let me open connections to them
Conns: rcmgr.DefaultLimit,
ConnsOutbound: rcmgr.DefaultLimit,
// No inbound streams from this peer
StreamsInbound: rcmgr.BlockAllLimit,
// And let me open unlimited (by me) outbound streams (the peer may have their own limits on me)
StreamsOutbound: rcmgr.Unlimited,
},
},
}
jsonBytes, _ := json.Marshal(&cfg)
// string(jsonBytes)
// {
// "System": {
// "StreamsOutbound": "unlimited"
// },
// "Peer": {
// "QmVvtzcZgCkMnSFf2dnrBPXrWuNFWNM9J3MpZQCvWPuVZf": {
// "StreamsInbound": "blockAll",
// "StreamsOutbound": "unlimited",
// "ConnsInbound": "blockAll"
// }
// }
// }This will omit defaults from the JSON output. It will also serialize the blockAll, and unlimited values explicitly.
The Memory field is serialized as a string to workaround the JSON limitation
of 32 bit integers (Memory is an int64).
Perhaps the most fundamental resource is memory, and in particular buffers used for network operations. The system must provide an interface for components to reserve memory that accounts for buffers (and possibly other live objects), which is scoped within the component. Before a new buffer is allocated, the component should try a memory reservation, which can fail if the resource limit is exceeded. It is then up to the component to react to the error condition, depending on the situation. For example, a muxer failing to grow a buffer in response to a window change should simply retain the old buffer and operate at perhaps degraded performance.
File descriptors are an important resource that uses memory (and computational time) at the system level. They are also a scarce resource, as typically (unless the user explicitly intervenes) they are constrained by the system. Exhaustion of file descriptors may render the application incapable of operating (e.g., because it is unable to open a file). This is important for libp2p because most operating systems represent sockets as file descriptors.
Connections are a higher-level concept endemic to libp2p; in order to communicate with another peer, a connection must first be established. Connections are an important resource in libp2p, as they consume memory, goroutines, and possibly file descriptors.
We distinguish between inbound and outbound connections, as the former are initiated by remote peers and consume resources in response to network events and thus need to be tightly controlled in order to protect the application from overload or attack. Outbound connections are typically initiated by the application's volition and don't need to be controlled as tightly. However, outbound connections still consume resources and may be initiated in response to network events because of (potentially faulty) application logic, so they still need to be constrained.
Streams are the fundamental object of interaction in libp2p; all protocol interactions happen through a stream that goes over some connection. Streams are a fundamental resource in libp2p, as they consume memory and goroutines at all levels of the stack.
Streams always belong to a peer, specify a protocol and they may belong to some service in the system. Hence, this suggests that apart from global limits, we can constrain stream usage at finer granularity, at the protocol and service level.
Once again, we distinguish between inbound and outbound streams. Inbound streams are initiated by remote peers and consume resources in response to network events; controlling inbound stream usage is again paramount for protecting the system from overload or attack. Outbound streams are normally initiated by the application or some service in the system in order to effect some protocol interaction. However, they can also be initiated in response to network events because of application or service logic, so we still need to constrain them.
The Resource Manager is based on the concept of resource scopes. Resource Scopes account for resource usage that is temporally delimited for the span of the scope. Resource Scopes conceptually form a DAG, providing us with a mechanism to enforce multiresolution resource accounting. Downstream resource usage is aggregated at scopes higher up the graph.
The following diagram depicts the canonical scope graph:
System
+------------> Transient.............+................+
| . .
+------------> Service------------- . ----------+ .
| . | .
+-------------> Protocol----------- . ----------+ .
| . | .
+-------------->* Peer \/ | .
+------------> Connection | .
| \/ \/
+---------------------------> Stream
The system scope is the top level scope that accounts for global resource usage at all levels of the system. This scope nests and constrains all other scopes and institutes global hard limits.
The transient scope accounts for resources that are in the process of full establishment. For instance, a new connection prior to the handshake does not belong to any peer, but it still needs to be constrained as this opens an avenue for attacks in transient resource usage. Similarly, a stream that has not negotiated a protocol yet is constrained by the transient scope.
The transient scope effectively represents a DMZ (DeMilitarized Zone), where resource usage can be accounted for connections and streams that are not fully established.
Same as the normal system scope above, but is used if the normal system scope is already at its limits and the resource is from an allowlisted peer. See Allowlisting multiaddrs to mitigate eclipse attacks see for more information.
Same as the normal transient scope above, but is used if the normal transient scope is already at its limits and the resource is from an allowlisted peer. See Allowlisting multiaddrs to mitigate eclipse attacks see for more information.
The system is typically organized across services, which may be ambient and provide basic functionality to the system (e.g. identify, autonat, relay, etc). Alternatively, services may be explicitly instantiated by the application, and provide core components of its functionality (e.g. pubsub, the DHT, etc).
Services are logical groupings of streams that implement protocol flow and may additionally consume resources such as memory. Services typically have at least one stream handler, so they are subject to inbound stream creation and resource usage in response to network events. As such, the system explicitly models them allowing for isolated resource usage that can be tuned by the user.
Protocol Scopes account for resources at the protocol level. They are an intermediate resource scope which can constrain streams which may not have a service associated or for resource control within a service. It also provides an opportunity for system operators to explicitly restrict specific protocols.
For instance, a service that is not aware of the resource manager and has not been ported to mark its streams, may still gain limits transparently without any programmer intervention. Furthermore, the protocol scope can constrain resource usage for services that implement multiple protocols for the sake of backwards compatibility. A tighter limit in some older protocol can protect the application from resource consumption caused by legacy clients or potential attacks.
For a concrete example, consider pubsub with the gossipsub router: the service also understands the floodsub protocol for backwards compatibility and support for unsophisticated clients that are lagging in the implementation effort. By specifying a lower limit for the floodsub protocol, we can can constrain the service level for legacy clients using an inefficient protocol.
The peer scope accounts for resource usage by an individual peer. This constrains connections and streams and limits the blast radius of resource consumption by a single remote peer.
This ensures that no single peer can use more resources than allowed by the peer limits. Every peer has a default limit, but the programmer may raise (or lower) limits for specific peers.
The connection scope is delimited to the duration of a connection and constrains resource usage by a single connection. The scope is a leaf in the DAG, with a span that begins when a connection is established and ends when the connection is closed. Its resources are aggregated to the resource usage of a peer.
The stream scope is delimited to the duration of a stream, and constrains resource usage by a single stream. This scope is also a leaf in the DAG, with span that begins when a stream is created and ends when the stream is closed. Its resources are aggregated to the resource usage of a peer, and constrained by a service and protocol scope.
User transaction scopes can be created as a child of any extant resource scope, and provide the programmer with a delimited scope for easy resource accounting. Transactions may form a tree that is rooted to some canonical scope in the scope DAG.
For instance, a programmer may create a transaction scope within a service that accounts for some control flow delimited resource usage. Similarly, a programmer may create a transaction scope for some interaction within a stream, e.g. a Request/Response interaction that uses a buffer.
Each resource scope has an associated limit object, which designates limits for all basic resources. The limit is checked every time some resource is reserved and provides the system with an opportunity to constrain resource usage.
There are separate limits for each class of scope, allowing for multiresolution and aggregate resource accounting. As such, we have limits for the system and transient scopes, default and specific limits for services, protocols, and peers, and limits for connections and streams.
When building software that is supposed to run on many different kind of machines, with various memory and CPU configurations, it is desirable to have limits that scale with the size of the machine.
This is done using the ScalingLimitConfig. For every scope, this configuration
struct defines the absolutely bare minimum limits, and an (optional) increase of
these limits, which will be applied on nodes that have sufficient memory.
A ScalingLimitConfig can be converted into a ConcreteLimitConfig (which can then be
used to initialize a fixed limiter with NewFixedLimiter) by calling the Scale method.
The Scale method takes two parameters: the amount of memory and the number of file
descriptors that an application is willing to dedicate to libp2p.
These amounts will differ between use cases. A blockchain node running on a dedicated server might have a lot of memory, and dedicate 1/4 of that memory to libp2p. On the other end of the spectrum, a desktop companion application running as a background task on a consumer laptop will probably dedicate significantly less than 1/4 of its system memory to libp2p.
For convenience, the ScalingLimitConfig also provides an AutoScale method,
which determines the amount of memory and file descriptors available on the
system, and dedicates up to 1/8 of the memory and 1/2 of the file descriptors to
libp2p.
For example, one might set:
var scalingLimits = ScalingLimitConfig{
SystemBaseLimit: BaseLimit{
ConnsInbound: 64,
ConnsOutbound: 128,
Conns: 128,
StreamsInbound: 512,
StreamsOutbound: 1024,
Streams: 1024,
Memory: 128 << 20,
FD: 256,
},
SystemLimitIncrease: BaseLimitIncrease{
ConnsInbound: 32,
ConnsOutbound: 64,
Conns: 64,
StreamsInbound: 256,
StreamsOutbound: 512,
Streams: 512,
Memory: 256 << 20,
FDFraction: 1,
},
}The base limit (SystemBaseLimit) here is the minimum configuration that any
node will have, no matter how little memory it possesses. For every GB of memory
passed into the Scale method, an increase of (SystemLimitIncrease) is added.
For Example, calling Scale with 4 GB of memory will result in a limit of 384 for
Conns (128 + 4*64).
The FDFraction defines how many of the file descriptors are allocated to this
scope. In the example above, when called with a file descriptor value of 1000,
this would result in a limit of 1000 (1000 * 1) file descriptors for the system
scope. See TestReadmeExample in limit_test.go.
Note that we only showed the configuration for the system scope here, equivalent configuration options apply to all other scopes as well.
By default the resource manager ships with some reasonable scaling limits and
makes a reasonable guess at how much system memory you want to dedicate to the
go-libp2p process. For the default definitions see DefaultLimits and
ScalingLimitConfig.AutoScale().
If the defaults seem mostly okay, but you want to adjust one facet you can
simply copy the default struct object and update the field you want to change. You can
apply changes to a BaseLimit, BaseLimitIncrease, and ConcreteLimitConfig with
.Apply.
Example
// An example on how to tweak the default limits
tweakedDefaults := DefaultLimits
tweakedDefaults.ProtocolBaseLimit.Streams = 1024
tweakedDefaults.ProtocolBaseLimit.StreamsInbound = 512
tweakedDefaults.ProtocolBaseLimit.StreamsOutbound = 512
Once you've set your limits and monitoring (see Monitoring below)
you can now tune your limits better. The rcmgr_blocked_resources metric will
tell you what was blocked and for what scope. If you see a steady stream of
these blocked requests it means your resource limits are too low for your usage.
If you see a rare sudden spike, this is okay and it means the resource manager
protected you from some anomaly.
Sometimes disabling all limits is useful when you want to see how much
resources you use during normal operation. You can then use this information to
define your initial limits. Disable the limits by using InfiniteLimits.
These errors occur whenever a limit is hit. For example, you'll get this error if you are at your limit for the number of streams you can have, and you try to open one more.
Example Log:
2022-08-12T15:49:35.459-0700 DEBUG rcmgr go-libp2p-resource-manager@v0.5.3/scope.go:541 blocked connection from constraining edge {"scope": "conn-19667", "edge": "system", "direction": "Inbound", "usefd": false, "current": 100, "attempted": 1, "limit": 100, "stat": {"NumStreamsInbound":28,"NumStreamsOutbound":66,"NumConnsInbound":37,"NumConnsOutbound":63,"NumFD":33,"Memory":8687616}, "error": "system: cannot reserve connection: resource limit exceeded"}
The log line above is an example log line that gets emitted if you enable debug
logging in the resource manager. You can do this by setting the environment
variable GOLOG_LOG_LEVEL="rcmgr=debug". By default only the error is
returned to the caller, and nothing is logged by the resource manager itself.
The log line message (and returned error) will tell you which resource limit was hit (connection in the log above) and what blocked it (in this case it was the system scope that blocked it). The log will also include some more information about the current usage of the resources. In the example log above, there is a limit of 100 connections, and you can see that we have 37 inbound connections and 63 outbound connections. We've reached the limit and the resource manager will block any further connections.
The next step in debugging is seeing if this is a recurring problem or just a transient error. If it's a transient error it's okay to ignore it since the resource manager was doing its job in keeping resource usage under the limit. If it's recurring then you should understand what's causing you to hit these limits and either refactor your application or raise the limits.
To check if it's a recurring problem you can count the number of times you've
seen the "resource limit exceeded" error over time. You can also check the
rcmgr_blocked_resources metric to see how many times the resource manager has
blocked a resource over time.
If the resource is blocked by a protocol-level scope, take a look at the various
resource usages in the metrics. For example, if you run into a new stream being blocked,
you can check the
rcmgr_streams metric and the "Streams by protocol" graph in the Grafana
dashboard (assuming you've set that up or something similar – see
Monitoring) to understand the usage pattern of that specific
protocol. This can help answer questions such as: "Am I constantly around my
limit?", "Does it make sense to raise my limit?", "Are there any patterns around
hitting this limit?", and "should I refactor my protocol implementation?"
Once you have limits set, you'll want to monitor to see if you're running into your limits often. This could be a sign that you need to raise your limits (your process is more intensive than you originally thought) or that you need to fix something in your application (surely you don't need over 1000 streams?).
There are Prometheus metrics that can be hooked up to the resource manager. See
obs/stats_test.go for an example on how to enable this, and DefaultViews in
stats.go for recommended views. These metrics can be hooked up to Prometheus
or any other platform that can scrape a prometheus endpoint.
There is also an included Grafana dashboard to help kickstart your observability into the resource manager. Find more information about it at here.
If you have a set of trusted peers and IP addresses, you can use the resource manager's Allowlist to protect yourself from eclipse attacks. The set of peers in the allowlist will have their own limits in case the normal limits are reached. This means you will always be able to connect to these trusted peers even if you've already reached your system limits.
Look at WithAllowlistedMultiaddrs and its example in the GoDoc to learn more.
go-libp2p already includes a connection
manager,
so what's the difference between the ConnManager and the ResourceManager?
ConnManager:
- Configured with a low and high watermark number of connections.
- Attempts to maintain the number of connections between the low and high markers.
- Connections can be given metadata and weight (e.g. a hole punched connection is more valuable than a connection to a publicly addressable endpoint since it took more effort to make the hole punched connection).
- The ConnManager will trim connections once the high watermark is reached. and trim down to the low watermark.
- Won't block adding another connection above the high watermark, but will trigger the trim mentioned above.
- Can trim and prioritize connections with custom logic.
- No concept of scopes (like the resource manager).
Resource Manager:
- Configured with limits on the number of outgoing and incoming connections at different resource scopes.
- Will block adding any more connections if any of the scope-specific limits would be exceeded.
The natural question when comparing these two managers is "how do the watermarks and limits interact with each other?". The short answer is that they don't know about each other. This can lead to some surprising subtleties, such as the trimming never happening because the resource manager's limit is lower than the high watermark. This is confusing, and we'd like to fix it. The issue is captured in go-libp2p#1640.
When configuring the resource manager and connection manager, you should set the limits in the resource manager as your hard limits that you would never want to go over, and set the low/high watermarks as the range at which your application works best.
Here we consider some concrete examples that can elucidate the abstract design as described so far.
Let's consider a stream and the limits that apply to it.
When the stream scope is first opened, it is created by calling
ResourceManager.OpenStream.
Initially the stream is constrained by:
- the system scope, where global hard limits apply.
- the transient scope, where unnegotiated streams live.
- the peer scope, where the limits for the peer at the other end of the stream apply.
Once the protocol has been negotiated, the protocol is set by calling
StreamManagementScope.SetProtocol. The constraint from the
transient scope is removed and the stream is now constrained by the
protocol instead.
More specifically, the following constraints apply:
- the system scope, where global hard limits apply.
- the peer scope, where the limits for the peer at the other end of the stream apply.
- the protocol scope, where the limits of the specific protocol used apply.
The existence of the protocol limit allows us to implicitly constrain
streams for services that have not been ported to the resource manager
yet. Once the programmer attaches a stream to a service by calling
StreamScope.SetService, the stream resources are aggregated and constrained
by the service scope in addition to its protocol scope.
More specifically the following constraints apply:
- the system scope, where global hard limits apply.
- the peer scope, where the limits for the peer at the other end of the stream apply.
- the service scope, where the limits of the specific service owning the stream apply.
- the protocol scope, where the limits of the specific protocol for the stream apply.
The resource transfer that happens in the SetProtocol and SetService
gives the opportunity to the resource manager to gate the streams. If
the transfer results in exceeding the scope limits, then a error
indicating "resource limit exceeded" is returned. The wrapped error
includes the name of the scope rejecting the resource acquisition to
aid understanding of applicable limits. Note that the (wrapped) error
implements net.Error and is marked as temporary, so that the
programmer can handle by backoff retry.
- The package only exports a constructor for the resource manager and basic types for defining limits. Internals are not exposed.
- Internally, there is a resources object that is embedded in every scope and implements resource accounting.
- There is a single implementation of a generic resource scope, that provides all necessary interface methods.
- There are concrete types for all canonical scopes, embedding a pointer to a generic resource scope.
- Peer and Protocol scopes, which may be created in response to network events, are periodically garbage collected.
- The Resource Manager must account for basic resource usage at all levels of the stack, from the internals to application components that use the network facilities of libp2p.
- Basic resources include memory, streams, connections, and file descriptors. These account for both space and time used by the stack, as each resource has a direct effect on the system availability and performance.
- The design must support seamless integration for user applications, which should reap the benefits of resource management without any changes. That is, existing applications should be oblivious of the resource manager and transparently obtain limits which protects it from resource exhaustion and OOM conditions.
- At the same time, the design must support opt-in resource usage accounting for applications that want to explicitly utilize the facilities of the system to inform about and constrain their own resource usage.
- The design must allow the user to set their own limits, which can be static (fixed) or dynamic.