Performance issues for production usage about olric HOT 4 OPEN

P50 is 0.5ms, P90 is 1.2ms, P95 is 2ms, P99 is 4ms

I think these numbers are pretty good. Olric is implement in Go.

If we create a worker pool of goroutines with the same client, with each worker having its own DMap with the same name. Then use the worker pool to process incoming requests rather than using a single DMap, will it help improve performance?

I have never tried such a thing before but increasing amount of context switches may reduce the performance at some point.

Won't owner always win in this case, or do I miss sth? It would be great if we can have a option to read from owner/primary only.

It depends what happened in the cluster before you run the request. If you are adding and removing the nodes frequently, you may encounter such anomalies. Only the owner node has read-write right on the partitions but it apply the LWW policy to return the most up to date result.

Olric is an AP store, that means Olric always chooses availability over consistency.

IIUC, the difference is that when owner is down, LWW can still read from backup, while with this new option, we need to wait for new node coming up to propagate the data from backup, during which the data is unavailable?

When a node goes down, a new partition owner will be assigned immediately and start processing the incoming requests. There is no active anti-entropy mechanism in Olric. It only tries to read keys from members(using the routing table, based on a consistent hash algorithm) and apply read-repair if it's enabled.

Also the calls to dm.lookupOnOwners and dm.lookupOnReplicas in getOnCluster are sequential. Can it be parallelized to speed up Get?

It may improve performance in some cases but I think the increasing amount of parallel network calls may decrease the overall performance for some workloads. This should be carefully designed and tested.

Besides, condition dm.s.config.ReadQuorum >= config.MinimumReplicaCount is always true, so dm.lookupOnReplicas will always be called.

Yeah, this enables the LWW implicitly.

from olric.

Hey @zhp007,

The setup, configuration, and results seem normal to me.

P99 set is 4.5ms, P99 get is 6ms. It is much higher than we expected.

What are the results for other percentiles? The Go runtime causes latency fluctuations. It might be a good idea to play with the GC parameters.

Will create more than one EmbeddedClient and/or DMap in each server help?

The EmbeddecClient/DMap implementation is thread-safe. With the same client instance, you can create any number of goroutines. Two CPUs should be good enough to get a rough idea of the performance characteristics.

Any other config settings or tunings we need to care about?

Currently, there are no other configuration options to improve performance. Still, you have two replicas, and Olric is trying to fetch all accessible values from the cluster before returning the result to the client. This is called the Last Write Wins(LWW) policy. It compares the timestamps in the returned DMap entries and the most up-to-date result wins. It's not possible to turn off this behavior explicitly. We can quickly implement it, but disabling LWW decreases the consistency.

See this:

ReadQuorum is one by default. We can add a boolean flag to turn off fetching the values from replicas but this will hurt the consistency.

The other thing to know is that if you request a key/value pair that does not belong to the node, the node finds the partition owner, fetches the pair from the owner, and returns it to the client. So, there is no redirection message in the protocol. It works as a reverse proxy.

See this:

from olric.

@buraksezer Thanks for quick reply!

What are the results for other percentiles?

Set P50 is 0.8ms, P90 is 1.7ms, P95 is 2.3ms
Get P50 is 1.5ms, P90 is 2.4ms, P95 is 3.1ms

We also tried ReplicaCount=1, there is no change on Set, while Get has better performance:
P50 is 0.5ms, P90 is 1.2ms, P95 is 2ms, P99 is 4ms

With the same client instance, you can create any number of goroutines.

If we create a worker pool of goroutines with the same client, with each worker having its own DMap with the same name. Then use the worker pool to process incoming requests rather than using a single DMap, will it help improve performance?

Olric is trying to fetch all accessible values from the cluster before returning the result to the client. This is called the Last Write Wins(LWW) policy. It compares the timestamps in the returned DMap entries and the most up-to-date result wins.

Won't owner always win in this case, or do I miss sth? It would be great if we can have a option to read from owner/primary only.
IIUC, the difference is that when owner is down, LWW can still read from backup, while with this new option, we need to wait for new node coming up to propagate the data from backup, during which the data is unavailable?

Also the calls to dm.lookupOnOwners and dm.lookupOnReplicas in getOnCluster are sequential. Can it be parallelized to speed up Get?

Besides, condition dm.s.config.ReadQuorum >= config.MinimumReplicaCount is always true, so dm.lookupOnReplicas will always be called.

The other thing to know is that if you request a key/value pair that does not belong to the node, the node finds the partition owner, fetches the pair from the owner, and returns it to the client.

Yes, understand that one additional hopping and data transfer can increase overhead/latency.

from olric.