From: Matt Corallo Date: Sat, 9 Dec 2023 21:22:52 +0000 (+0000) Subject: Layout channel info to ensure routing uses cache lines well X-Git-Tag: v0.0.124-beta~45^2 X-Git-Url: http://git.bitcoin.ninja/?a=commitdiff_plain;h=f689e01b358dfa111596721a9a84f808133cd4f9;p=rust-lightning Layout channel info to ensure routing uses cache lines well Because we scan per-channel information in the hot inner loop of our routefinding immediately after looking a channel up in a `HashMap`, we end up spending a nontrivial portion of our routefinding time waiting on memory to be read in. While there is only so much we can do about that, ensuring the channel information that we care about is sitting on one or adjacent cache lines avoids paying that penalty twice. Thus, here we manually lay out `ChannelInfo` and `ChannelUpdateInfo` and set them to 128b and 32b alignment, respectively. This wastes some space in memory in our network graph, but improves routing performance in return. --- diff --git a/lightning/src/routing/gossip.rs b/lightning/src/routing/gossip.rs index b0117cd68..358bf7095 100644 --- a/lightning/src/routing/gossip.rs +++ b/lightning/src/routing/gossip.rs @@ -795,22 +795,32 @@ where } } +// Fetching values from this struct is very performance sensitive during routefinding. Thus, we +// want to ensure that all of the fields we care about (all of them except `last_update_message`) +// sit on the same cache line. +// +// We do this by using `repr(C)`, which forces the struct to be laid out in memory the way we write +// it (ensuring `last_update_message` hangs off the end and no fields are reordered after it), and +// `align(32)`, ensuring the struct starts either at the start, or in the middle, of a 64b x86-64 +// cache line. This ensures the beginning fields (which are 31 bytes) all sit in the same cache +// line. +#[repr(C, align(32))] #[derive(Clone, Debug, PartialEq, Eq)] /// Details about one direction of a channel as received within a [`ChannelUpdate`]. pub struct ChannelUpdateInfo { - /// When the last update to the channel direction was issued. - /// Value is opaque, as set in the announcement. - pub last_update: u32, - /// Whether the channel can be currently used for payments (in this one direction). - pub enabled: bool, - /// The difference in CLTV values that you must have when routing through this channel. - pub cltv_expiry_delta: u16, /// The minimum value, which must be relayed to the next hop via the channel pub htlc_minimum_msat: u64, /// The maximum value which may be relayed to the next hop via the channel. pub htlc_maximum_msat: u64, /// Fees charged when the channel is used for routing pub fees: RoutingFees, + /// When the last update to the channel direction was issued. + /// Value is opaque, as set in the announcement. + pub last_update: u32, + /// The difference in CLTV values that you must have when routing through this channel. + pub cltv_expiry_delta: u16, + /// Whether the channel can be currently used for payments (in this one direction). + pub enabled: bool, /// Most recent update for the channel received from the network /// Mostly redundant with the data we store in fields explicitly. /// Everything else is useful only for sending out for initial routing sync. @@ -878,22 +888,46 @@ impl Readable for ChannelUpdateInfo { } } +// Fetching values from this struct is very performance sensitive during routefinding. Thus, we +// want to ensure that all of the fields we care about (all of them except `last_update_message` +// and `announcement_received_time`) sit on the same cache line. +// +// Sadly, this is not possible, however we can still do okay - all of the fields before +// `one_to_two` and `two_to_one` are just under 128 bytes long, so we can ensure they sit on +// adjacent cache lines (which are generally fetched together in x86_64 processors). +// +// This leaves only the two directional channel info structs on separate cache lines. +// +// We accomplish this using `repr(C)`, which forces the struct to be laid out in memory the way we +// write it (ensuring the fields we care about are at the start of the struct) and `align(128)`, +// ensuring the struct starts at the beginning of two adjacent 64b x86-64 cache lines. +#[repr(align(128), C)] #[derive(Clone, Debug, Eq)] /// Details about a channel (both directions). /// Received within a channel announcement. pub struct ChannelInfo { /// Protocol features of a channel communicated during its announcement pub features: ChannelFeatures, + /// Source node of the first direction of a channel pub node_one: NodeId, - /// Details about the first direction of a channel - pub one_to_two: Option, + /// Source node of the second direction of a channel pub node_two: NodeId, - /// Details about the second direction of a channel - pub two_to_one: Option, + + /// The [`NodeInfo::node_counter`] of the node pointed to by [`Self::node_one`]. + pub(crate) node_one_counter: u32, + /// The [`NodeInfo::node_counter`] of the node pointed to by [`Self::node_two`]. + pub(crate) node_two_counter: u32, + /// The channel capacity as seen on-chain, if chain lookup is available. pub capacity_sats: Option, + + /// Details about the first direction of a channel + pub one_to_two: Option, + /// Details about the second direction of a channel + pub two_to_one: Option, + /// An initial announcement of the channel /// Mostly redundant with the data we store in fields explicitly. /// Everything else is useful only for sending out for initial routing sync. @@ -903,11 +937,6 @@ pub struct ChannelInfo { /// (which we can probably assume we are - no-std environments probably won't have a full /// network graph in memory!). announcement_received_time: u64, - - /// The [`NodeInfo::node_counter`] of the node pointed to by [`Self::node_one`]. - pub(crate) node_one_counter: u32, - /// The [`NodeInfo::node_counter`] of the node pointed to by [`Self::node_two`]. - pub(crate) node_two_counter: u32, } impl PartialEq for ChannelInfo {