Hello All
I have read some of VAST Data blog about KV cashe and AI agent.
And several of tests in that articles use VAST NFS driver.
I understand that this is primarily used for multipath purposes, but is it possible to improve the KV cache hit rate by optimizing the mount options of the NFS driver?
As we are explaining VAST Data to customers, it would be helpful if we could describe such an effect, if applicable.
Best Regards
Hi @kodai, thanks for posting, will ask around and get a response to you!
Hi Kodai, thank you for your question!
Optimizing the NFS driver mount options will not improve the logical KV cache hit rate. This role is under the KV$ management system of your choice that is responsible to manage and map the kvcache. But, given the same storage capacity, thanks to vast data reduction that we can save around 30% of kv$ data, within the same storage capacity we can store more kv$ data, hence increase also kv$ hit rate.
vast nfs multi path improve KV cache performance:
To get the best performance, we recommend using NFSv3 with our multipath driver and tuning the read-ahead configurations as documented here
# Example NFSv3 Multipath Mount Command
mount -o vers=3,nconnect=32,pconnect=4,mdconnect=2,spread_reads,spread_writes,noextend,remoteports=dns ``vast.example.com``:/export /mnt/vast
3. Application-Level Tuning
Finally, there are application-level configurations that also greatly affect KV cache offloading performance. For example, if the customer is using lmcache, optimizing parameters in the YAML file—such as increasing chunk_size and local_cpu_size, local_disk_path_sharding ,.. Can make a big difference depending on their specific deployment scale.
I would be happy to jump on a call to discuss any deployment-specific questions or help you position this with the customer!
please look also at the nfs multi path guide to get started
Hi @anat.heilper-san @ram -san
Thank you very much for answering the question.
I understand it. I will use this theory for next disscution.
Best regards