High-level summary​

During the past two weeks, the consensus team started documenting the implementation of the UTxO HD feature and continued developing tests for it. As part of our work on UTxO HD, we improved the Haskell support for LMDB. We also spent time working on the LSM tree prototype, and designed a parameter tuning algorithm for it. Regarding our work on Genesis, our investigation of the "plateaus" pointed at the TICKF slowdown on era boundaries as culprit. This led us to developing a caching strategy that will not only remove the aforementioned "plateaus", but can help alleviating the growing block production delay on epoch switch. We also helped reviewing the block forge credential hotswap feature, which is intended for use in the adoption of P2P.

We also worked on paying technical debt and fostering collaboration. In particular, we improved the io-sim framework, which is crucial for testing and simulating Cardano components. We also removed thunks that appeared on era translations, and improved our diffusion pipelining feature. We are working on a presentation for explaining Praos and Genesis.

High-level status report​

• Finish the UTxO HD prototype: in progress.
  • We added documentation for this feature.
  • We developed the second version of the mempool tests.
  • We fixed benchmarks that were inflating the speedup we observed in the anti-diff implementation of sequences of differences. Speedups are now in the range of [3.33, 4.75], which remain significant.
  • We continued improving Haskell LMDB support.
  • We finished implementing a "parameter tuning algorithm" for the LSM tree prototype. This enables us to run experiments to check the correctness of the algorithm.
• Genesis: in progress.
  • Work investigating the "plateaus" in the ChainSync jumping prototype pointed to the TICKF slowdown on era boundaries as culprit.
• Tech debt:
  • We improved the capabilities of our io-sim library, which is crucial for testing and simulating Cardano components.
  • We removed thunks from epoch translations in the ledger.
  • We added Linux CI support for lmdb-simple.
  • We got pending diffusion pipelining improvements merged.
• Fostering collaboration:
  • We are working on a explanation of Praos and Genesis protocols.
• Support:
  • Investigation of CSJ "plateaus" led us to developing a caching strategy for TICKF that will not only remove these "plateaus", but can help alleviating the growing block production delay on epoch switch.
  • We reviewed the block forge credential hotswapping feature which is intended for use in the adoption of P2P.

Workstreams​

Finish the UTxO HD prototype​

We merged PR #4060, which adds a report documenting the UTxO HD feature, and puts emphasis in explaining how the mempool works in combination with UTxO HD.

We opened a draft PR with the second iteration of the property tests for the mempool (#4076).

We fixed the Arbitrary instances for keys and values in DiffSeq benchmarks (#4143). The problem was that we were testing with mostly small values, which artificially boosted the performance gains we saw on benhcmarks. Speedups are now in the range of [3.33, 4.75] across the different configurations.

Backing store property tests​

We focused on incorporating feedback on the monadic cursor API PR (#1). This required us to make small tweaks to quickcheck-lockstep to test the new API. We also updated the backing store property tests to use the new version of the monadic cursor API.

LSM tree implementation​

We worked on the LSM tree prototype. In particular: finished implementing a "parameter tuning algorithm" that adapts the LSM tree design based on factors like:

• workload
• machine specs,
• and characteristics of the data being stored.

We are now running experiments to gather results and cross-reference them with existing experimental results from the LSM tree paper to see if the algorithm is working correctly.

Benchmarking the CSJ prototype​

We focused on investigating the "plateaus" in the ChainSync tip, which turned out to be due to the TICKF bug which we previously were only aware of in the context of the long forging times near epoch boundaries. For the most drastic patch by @nfrisby to speed up TICKF, full sync is speeding up by 7%.

The following plot shows that by caching the TICKF the ChainSync tip and the VolatileDB tip progress at the same rate.

The plot below shows the speedup observed by caching the TICKF rule wrt the baseline.

Technical debt​

After addressing the PR comments, we merged PR #16, which implements the MonadCatch instance for STM. This extends the capability of our io-sim library, which is crucial for testing and simulating Cardano components PR #16 closed #1461. This new feature was published as version 0.4.0.0 of io-sim.

We continued with our work fixing the NoThunk errors required for enabling nightly tests, with the help of TVarInvariant checks in strict-stm and nothunks libraries. We proposed fixes in cardano-ledger that took care of thunks that appeared in era translations (#3143). The fixes will be integrated back into consensus when cardano-ledger approves and publish the changes introduced in #3143.

We added CI support for lmdb-simple (#2). We currently test the build on a Linux environment only.

We got pending diffusion pipelining PRs (#3857, #3860, #3856) merged, after rebasing and addressing feedback.

Fostering collaboration​

@nfrisby finished a visualisation tool and outlined scripts for the Praos and Genesis explanation presentations. The idea is to produce a video that gives an overview of these protocols.

Support​

We started working on caching the computation of the TICKF rule (#4054), since this was blocking our benchmarking work for Genesis. In addition, this issue has the Cardano community quite concerned, so we are hoping the work done in caching the computation of the TICKF rule can help alleviating the growing block production delay on epoch switch.

We reviewed the block forge credential hotswapping PR #3800 from the networking team, which is intended for use in the adoption of P2P.

High-level summary​

During the past two weeks, the consensus team continued its work on testing the UTxO HD prototype. We completed the era-transition and backing store tests, and the mempool tests are advancing at a steady pace. Regarding our work in the Genesis design, we continued our collaboration with the research and networking teams, and we continue investigating strategies for making the chain-sync jumping prototype faster.

High-level status report​

• Finish the UTxO HD prototype: on track.
  • We worked on state-machine tests for the mempool, and spotted potential bugs in the implementation. Investigation is ongoing.
  • We have a set of property tests for the backing store. We still need to incorporate the improvements to the LMDB cursor API that these tests made possible.
  • We merged the era-transition tests PR.
• Genesis: on track.
  • Design work around Genesis continues in collaboration with researchers and the networking team.
  • We continued trying to improve the performance of the chain-sync jumping prototype. We gained additional insight on which parameters to tweak next. In spite of the baseline still being faster, the current prototype already achieves a significant speedup when compared to the naive approach of simply running full chain-sync with all peers.
• Tech debt: on track.
  • We clarified a common source of confusion around VRF tie-breaking and cross-era chain selection.

Workstreams​

Finish the UTxO HD prototype​

We continued working on property-tests for the UTxO HD prototype. In particular we merged the era-transition tests PR.

Backing store property tests​

The backing store property tests PR has been reviewed. The next steps are:

• Improve error handling and command generation.
• Add coverage testing to check that we are not failing to cover interesting test cases.

The monadic cursor API went through its first review round. The API is in a relatively stable state. This PR also unifies the cborg and serialise-based interfaces to LMDB operations. The next steps are:

• Write quickcheck-dynamic state-machine tests for this API.
• Adapt the changes in the serialisation interface in the backing store property tests. This will involve adding boilerplate code in consensus to make up for the removal of the cborg-based interface.

LSM tree implementation​

We worked on the LSM tree prototype. In particular, we focused on tuning the LSM tree design to the different workloads that consensus has (eg syncing, normal node operation, etc).

Benchmarking the CSJ prototype​

Work on improving the chain-sync jumping performance is ongoing. In particular we compared the performance of different jump intervals, which, somewhat surprisingly, do not make a significant difference. In particular, we are seeing periodic "plateaus" where the chain-sync tip does not progress, but they are much longer for the prototype. Our hypothesis is that this seem to be due to a combination of the garbage collector (GC) pauses, and the actual time it takes the non-dynamo chain-sync peers to jump to the tip of the slot of the dynamo fragment.

In the coming weeks we will try to shorten these plateaus via a combination of tweaking GC options and less synchronisation in the CSJ governor.

The following plot shows the performance of the chain-sync jumping prototype using different jumping intervals. It compares the syncing progress by plotting the slots of adopted blocks against time. The baseline is still faster, however it is worth noting that the current prototype already achieves a significant speedup when compared to the naive approach of simply running full chain-sync with all peers.

The second plot shows the syncing progress sliced to a chosen ~5min interval, and includes, in addition to the slots of adopted blocks, the slots of the tip of the ChainSync fragment. This allows us to see how far ahead of the selected tip the CS dynamo is, i.e. how much room we have for BlockFetch not to get stalled. It shows periodic behaviour (due to the forecasting limit), and shows that the CS fragment tip is not progressing for significant periods ("plateaus").

Technical debt​

We clarified a common source of confusion around VRF tie-breaking and cross-era chain selection. This PR involved correcting potentially misleading names of VRF-related functions, and providing context for a particular VRF value is used for tie-breaking.

High-level summary​

During the past two weeks, the consensus team worked on adding property test for different aspects of the UTxO HD prototype: era transitions, mempool, and backing store. Thanks to these tests we were able to uncover a bug in the prototype. On the Genesis front, we benchmarked a different version of the ChainSync jumping prototype to try to improve its performance, but this did not result in any noticeable speedup.

High-level status report​

• Finish the UTxO HD prototype: on track.
  • We focused on increasing test coverage for the UTxO-HD prototype:
    • We started implementing Cadano-eras transition property-tests.
    • We started implementing state-machine property-tests for the mempool.
    • We merged the mempool rewrite.
    • We started working on state-machine tests for the backing store. This uncovered a bug in the range-read implementation of the LMDB backing store.
• Genesis: on track.
  • We benchmarked a version of the Genesis ChainSync Jumping prototype that spreads out the ChainSync updates over a longer period of time. This did not result in any noticeable speedup.
  • We investigated the overhead introduced by non-ChainSync components, but no conclusions could be drawn from the benchmarks we ran.

Workstreams​

Finish the UTxO HD prototype​

We focused on increasing test coverage for the UTxO HD prototype. We also merged the mempool rewrite.

Era transition property tests​

We started implementing Cardano era transition property tests, which are needed for making sure that the ledger tables get updated in the right way when we move from one era to the next. There are at the moment two important transitions.

• Byron to Shelley: where all the UTxO is transferred from in-memory Byron state (which has no tables) to the ledger tables of the Shelley state.
• Shelley to Allegra: where the AVVM addresses must be deleted.

We have tests for the Byron to Shelley transitions. We are working on adding the remaining ones.

Mempool state-machine tests​

We started implementing state-machine property tests for the mempool. The mempool is currently tested via pure property tests, and use a ledger state without tables. With the introduction of UTxO HD, testing the concurrent behavior of the mempool became of crucial importance (eg now we have to acquire locks to flush the backing store). In addition, we need to test a ledger state with tables. These needs led to the creation of a new set of property tests. In particular we aim to run parallel state-machine tests that exercise the mempool in a way similar to how the node would make use of it.

Backing store property tests​

We started working on state-machine tests for the backing store that UTxO HD uses. The property tests uncovered errors in the range-reads implementation of the LMDB backing store. To facilitate fixing this bug, we made changes to the Haskell LMDB bindings.

Benchmarking the CSJ prototype​

Prompted by previous benchmarks showing significant improvements in sync time by using more capabilities, we implemented a way to spread out the ChainSync updates over a larger period instead of firing them all at the same time. This didn't result in a noticeable speedup.

We also benchmarked the prototype with CSJ disabled (such that just the dynamo peer is running ChainSync, but e.g. BlockFetch still sees all peers) to rule out/confirm overhead by non-ChainSync (mainly BlockFetch) related components. This results in era-specific behavior (speed is like the prototype in Byron, but like the baseline in Shelley). This deserves a closer look in the future.

This diagram shows the respective syncing progress, starting at Genesis and continuing a good part into Shelley (with the dashed line indicating the Byron-to-Shelley transition).

• Red: baseline
• Green: CSJ prototype, 10 peers, jumps every 3000/f slots, jumps in clumps.
• Blue: like Green, jumps are spread out.
• Orange: variant with no jumping, to measure unrelated overhead.

High level summary​

During the past two weeks, the consensus team worked on improving the performance of the ChainSync jumping logic, which is needed for Genesis. We also rewrote the implementation of the mempool in the UTxO HD prototype which solved the issues that prevented us from running system level benchmarks. Also on the UTxO HD front, we have an improved implementation of the sequence-of-differences (a crucial piece of UTxO HD), and we also elaborated a test sign-off list for the UTxO HD feature.

Executive summary​

• With the latest implementation of ChainSync jumping we are closer to the baseline performance. In particular, the prototype seems to benefit from the extra concurrency provided by additional capabilities.
• We rewrote the implementation of the mempool in the UTxO HD prototype. This rewrite was required due to performance problems we observed when running the workbench. These performance problems prevented us from running system level benchmarks. The rewrite solved these issues. After the UTxO-HD: mempool rewrite PR is merged, we will contact the Benchmarking team so that they run the system level benchmarks.
• The implementation of sequences of differences based on anti-diffs was integrated into the UTxO HD prototype. It is pending review and we also need to run replay and syncing benchmarks to confirm that this will deliver a performance improvement, as observed in our micro-benchmarks.
• The UTxO HD prototype inspection resulted in a list of tests needed for consensus to consider the UTxO HD prototype as fully tested.

Additional information​

Genesis​

Benchmarking setup: 50MBit/s, 50ms latency

• Red: baseline
• Green: Current CSJ prototype, 10 peers, jumps every 3000/f slots.

As ChainSync Jumping involves many concurrent network operations at every jump, we tried to run the node with 6 instead of the default 2 capabilties.

• Orange: baseline with 6 capabilities
• Blue: CSJ prototype with 6 capabilities

This diagram shows the respective syncing progress, starting at Genesis and continuing a good part into Shelley (with the dashed line indicating the Byron-to-Shelley transition).

Further work includes whether we can tune the prototype to better handle few capabilities, or to adapt the default number of capabilities (potentially just while syncing).

• We proposed a fix for the performance degradation observed when running distributed multi-node benchmarks in the UTxO HD feature branch. While this fixed the problems observed when running local benchmarks, it broke the ThreadNet tests due to concurrency issues. Therefore, we think it is wise to start redesigning the UTxO HD mempool integration.
• We did several rounds of code review on the alternative implementation of diff-sequences required by the UTxO HD feature based on the idea of anti-diffs. This alternative implementation is close to being merged, and the next step is to integrate this to the UTxO HD branch, so that we can run ad-hoc replaying and syncing from scratch benchmarks and compare these with the baseline. The micro-benchmarks we elaborated for the alternative implementation show speedups of up to 4x, so we are optimistic about the performance of replaying and syncing from scratch benchmarks, however it is important to notice that due to the nature of UTxO HD we will still be slower than the baseline.
• The final draft of the Genesis implementation specification is ready for review.
• We implemented a prototype for the happy path of Genesis' ChainSync Jumping (CSJ). The prototype is slower than the baseline, however it is not the latest version of the prototype and the jump interval is very small.
• Work on integrating Conway has stopped since priorities have changed.
• We started work on benchmarking epoch-boundaries and epoch overhead pr-4014. To this end, we made use of a modified version of our db-analyser tool. We ran the new benchmarking setup using the Cardano mainnet chain, and we can see that block tick and application take substantially longer at epoch boundaries, although there are a couple of slots during an epoch in which these computations take more than normal. We notified the ledger team about these findings. We will use this modified version of db-analyser to investigate the epoch overhead.

Workstreams

UTxO HD​

• Spent quite some time investigating the root cause of the degradation in performance observed in the benchmarks. We run the make forge-stress benchmarks locally in order to debug this behavior.

  • Transaction batching doesn't make a notable difference in the outcome (considering we are using the in-memory backend).

  • The mempool batching implementation required asynchronous transaction validation which is a violation of the LocalTxSubmission protocol contract and therefore if we continued on that route, the impact would have been quite big.

  • The STM logic we implemented by using a TMVar for the mempool internal state was buggy and under certain circumstances it seemed to lock. Reverting the mempool internal state to be stored in a TVar seems to solve this problem.

  • The results we get after this change look almost identical to the ones from the baseline.

• The anti-diff prototype (PR #3997) has been reviewed and is close to being merged.

  • A follow-up issue (issue #4010) to integrate the anti-diff prototype in the various consensus packages was created. A first version of the integration exists, and all tests pass. A next step is to get some indication of the "real" performance gain by profiling db-analyser (or cardano-node).

Genesis​

• Final draft of the Genesis implementation specification, now up for review.

• Local benchmark setup for parameter tuning via the happy path ChainSync Jumping (CSJ) prototype (Issue 3987).

  • Context: Our Genesis design requires us to check in with a large (~20) number of servers periodically while syncing. These servers are offered jump requests via the ChainSync protocol (hence the name), which they can accept or decline. If a peer declines, the Genesis rule allows us to determine whether a node actually has a better chain.

  • The "happy path" is when no peer declines a jump. We want this to have close to no overhead compared to status quo, i.e. syncing without Genesis.

  • We implemented a prototype for this happy path, and are now starting to test in various configurations (number of peers, latency, bandwidth) to tune the performance of ChainSync jumping, i.e. how complicated our logic of choosing when to jump needs to be.

    Example:

  • Simulated connection: 50 MBit/s, 50ms latency

  • Jump interval: 3000 slots (on the low end, could be increased to up to 3k/f)

  • Red: baseline (1.35.3), one peer in topology file

  • Blue: Preliminary version of our prototype, with 10 peers.

    

    It is slower by about ~30%, but it is not the latest version of the prototype, and the jump interval is very small, making CSJ more of a bottleneck.

Technical debt​

• Fix flakiness in ChainDB QSM tests (PR 3990).