eeee75c5be
Don't use pinned memory when unnecessary ( #17832 )
...
Reports of excessive GPU memory usage and errors
from cudaHostRegister. There are some cases where pinning is
not required.
2021-06-14 16:10:04 +02:00
985280ec0b
excludes epoch-slots from nodes with unknown or different shred version ( #17899 )
...
Inspecting TDS gossip table shows that crds values of nodes with
different shred-versions are creeping in. Their epoch-slots are
accumulated in ClusterSlots causing bogus slots very far from current
root which are not purged and so cause ClusterSlots keep consuming more
memory:
https://github.com/solana-labs/solana/issues/17789
https://github.com/solana-labs/solana/issues/14366#issuecomment-769896036
https://github.com/solana-labs/solana/issues/14366#issuecomment-832754654
This commit updates ClusterInfo::get_epoch_slots, and discards entries
from nodes with unknown or different shred-version.
Follow up commits will patch gossip not to waste bandwidth and memory
over crds values of nodes with different shred-version.
2021-06-13 14:08:08 +00:00
cab30e2356
parallelizes gossip packets receiver with processing of requests ( #17647 )
...
Gossip packet processing is composed of two stages:
* The first is consuming packets from the socket, deserializing,
sanitizing and verifying them:
https://github.com/solana-labs/solana/blob/7f0349b29/gossip/src/cluster_info.rs#L2510-L2521
* The second is actually processing the requests/messages:
https://github.com/solana-labs/solana/blob/7f0349b29/gossip/src/cluster_info.rs#L2585-L2605
The former does not acquire any locks and so can be parallelized with
the later, allowing better pipelineing properties and smaller latency in
responding to gossip requests or propagating messages.
2021-06-07 18:36:06 +00:00
60b0a13444
writes epoch-slots to crds table synchronously ( #17719 )
...
epoch-slots may be overwritten before they are written to crds table:
https://github.com/solana-labs/solana/issues/17711
This commit writes new epoch-slots to crds table synchronously with
push_epoch_slots. The functions is still not thread-safe as commented in
the code, however currently only one threads is invoking this code.
2021-06-04 13:56:51 +00:00
be957f25c9
adds fallback logic if retransmit multicast fails ( #17714 )
...
In retransmit-stage, based on the packet.meta.seed and resulting
children/neighbors, each packet is sent to a different set of peers:
https://github.com/solana-labs/solana/blob/708bbcb00/core/src/retransmit_stage.rs#L421-L457
However, current code errors out as soon as a multicast call fails,
which will skip all the remaining packets:
https://github.com/solana-labs/solana/blob/708bbcb00/core/src/retransmit_stage.rs#L467-L470
This can exacerbate packets loss in turbine.
This commit:
* keeps iterating over retransmit packets for loop even if some
intermediate sends fail.
* adds a fallback to UdpSocket::send_to if multicast fails.
Recent discord chat:
https://discord.com/channels/428295358100013066/689412830075551748/849530845052403733
2021-06-04 12:16:37 +00:00
3a647c4bea
Rename ValidatorExit and move to sdk ( #17728 )
2021-06-04 03:06:13 +00:00
9a5330b7eb
Move gossip modules into solana-gossip crate ( #17352 )
...
* Move gossip modules to solana-gossip
* Update Protocol abi digest due to move
* Move gossip benches and hook up CI
* Remove unneeded Result entries
* Single use statements
2021-05-26 09:15:46 -06:00
sakridge
behzad nouri
Tyera Eulberg