- The write stage will output vector of entries
- Broadcast stage will create blobs out of the entries
- Helps reduce MIPS requirements for write stage
* Move register_entry_id() call out of write stage
- Write stage is MIPS intensive and has become a bottleneck for
TPU pipeline
- This will reduce the MIPS requirements for the stage
* Fix rust format issues
* budget and system contracts and verification
* contract check_id methods
* system call contract
* verify contract execution rules
* move system into its own file
* allocate before transfer for budget
* store error in budget context
* budget contract and tests without bank
* moved budget of of bank
* Tweak GCE scripts for higher node count
- Some validators were unable to rsync config from leader when
the node count was high (e.g. 25). Looks like the leader node was
getting more rsync requests in parallel than it count handle.
- This change staggers the validators bootup, and rsync time
* Address review comments
* Use multiple sockets for receiving blobs on validators
- The blobs that are broadcasted by leader or retransmitted by peer
validators are received on replicate_port
- Using reuse_addr/reuse_port, multiple sockets can be opened for
the same port
- This allows the kernel to queue data to user space app on multiple
socket queues, preventing over-running one queue
- This helps with reducing packets dropped due to queue over-runs
Fixes#1224
* Fixed failing tests
* Implemented recvmmsg() for UDP packets
- This change implements binding between libc API for recvmmsg()
- The function can receive multiple packets using one system call
Fixes#1141
* Added unit tests for recvmmsg()
* Added recv_mmsg() wrapper for non Linux OS
* Address review comments for recvmmsg()
* Remove unnecessary imports
* Moved target specific dependencies to the function
In the error case that i>0 (we have blobs to send)
we break out of the loop and do not push the allocated r
to the v array. We should recycle this blob, otherwise it
will be dropped.