Also, I didn't look at your network trace screen cap, but you should have zero TCP retransmissions if you set your initial TCP window send window small enough.
https://www.auvik.com/franklyit/blog/tcp-window-size/ On Fri, Jun 21, 2024, 07:53 Jeremy McMillan <[email protected]> wrote: > It could be network or persistent storage. What's the proportion of fast > to slow gets? > > > On Thu, Jun 20, 2024, 22:48 f cad <[email protected]> wrote: > >> here is a screenshot example >> [image: image.png] >> >> f cad <[email protected]> 于2024年6月21日周五 11:45写道： >> >>> Hello, community: >>> >>> I have a cluster ,with three nodes. >>> I have two cache, that AtomicityMode is TRANSACTIONAL and Backups number is >>> two and WriteSynchronizationMode is PRIMARY_SYNC >>> >>> >>> I use IgniteClientSpringTransactionManagerwith OPTIMISTIC transaction and >>> SERIALIZABLE concurrency mode. >>> >>> pseudocode like below >>> ignite.transactions().txStart >>> if(acahce.get(key)==null) { >>> aCahce.put(key,value) >>> bCahce.put(key,value) >>> } >>> tx.commit() >>> >>> Sometimes I find aCahce.get(key) costs 80ms ,Sometimes that costs only 5ms. >>> and three ignite nodes usage of cpu and io and memory all not high. >>> and client node usage of cpu and io and memory all not high. >>> but I use tcpdump to find that Between nodes, there are over 40 TCP >>> retransmissions per second. >>> So is this a network issue? >>> >>>
