An implementation of the libp2p network resource manager interface

Clone this repo:
  1. 2e798e3 release v0.3.0 (#23) by Gus Eggert · 8 months ago master v0.3.0
  2. 3690eb4 Merge pull request #21 from libp2p/feat/export-limiter-config by vyzo · 8 months ago
  3. 302a7ea feat: export basic limiter config by Gus Eggert · 8 months ago feat/export-limiter-config
  4. 573ebeb Merge pull request #20 from libp2p/fix/config-err-msgs by vyzo · 8 months ago
  5. 9c55f3b fix: typos in config err msgs by Gus Eggert · 8 months ago fix/config-err-msgs

The libp2p Network Resource Manager

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.

Design Considerations

  • 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 protect it from resource exhaustion and OOM conditions.
  • At the same time, the design must support opt-in resource usage accounting for applications who 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 its own limits, which can be static (fixed) or dynamic.

Basic Resources


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

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), most importantly 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 disinguish 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.

Resource Scopes

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:

  +------------> Transient.............+................+
  |                                    .                .
  +------------>  Service------------- . ----------+    .
  |                                    .           |    .
  +------------->  Protocol----------- . ----------+    .
  |                                    .           |    .
  +-------------->* Peer               \/          |    .
                     +------------> Connection     |    .
                     |                             \/   \/
                     +--------------------------->  Stream

The System Scope

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

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.

Service Scopes

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

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.

Peer Scopes

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.

Connection Scopes

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.

Stream Scopes

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

User transaction scopes can be created as a child of any extant resource scope, and provide the prgrammer 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 us 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.


Here we consider some concrete examples that can ellucidate the abstract design as described so far.

Stream Lifetime

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 protcol 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.

Default Limits

The provided default static limiters apply the following limits, where memoryCap is provided by the programmer, either as a fixed number (in bytes) or a fraction of total system memory:

 StreamsInbound:  4096,
 StreamsOutbound: 16384,
 ConnsInbound:    256,
 ConnsOutbound:   512,
 FD:              512,

 StreamsInbound:  128,
 StreamsOutbound: 512,
 ConnsInbound:    32,
 ConnsOutbound:   128,
 FD:              128,

 StreamsInbound:  1024,
 StreamsOutbound: 4096,

 StreamsInbound:  2048,
 StreamsOutbound: 8192,

 Memory:    memoryCap
 BaseLimit: DefaultSystemBaseLimit

 Memory:    memoryCap / 16
 BaseLimit: DefaultTransientBaseLimit

svc =
 Memory:    memoryCap / 2
 BaseLimit: DefaultServiceBaseLimit

 Memory:    memoryCap / 4
 BaseLimit: DefaultProtocolBaseLimit

 Memory:    memoryCap / 16
 BaseLimit: DefaultPeerBaseLimit

 Memory:    16 << 20,

 Memory:    16 << 20,

We also provide a dynamic limiter which uses the same base limits, but the memory limit is dynamically computed at each memory reservation check based on free memory.

Implementation Notes

  • 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.