asking for NVDIMM cache layer enablement

[gaowayne]

hello dear product owners, can I ask if this ATS server can enable NVDIMM as cache, it can works with pmem fsdax mode naturally.

[bryancall]

We support cache as a block device or a file. Depends on the OS and how it handles the NVDIMM.

[gaowayne]

@bryancall I see ATS support file or block device, I am sorry I am new to this, I did not see any tier at the block cache level. 1st layer is ram_cache, 2nd layer is block or file, if we put nvdimm or pmem into 2nd layer, they are small and expensive, better more hot data is in nvdimm or pmem, more cold data into HDD or QLC

[cukiernik]

@gaowayne, I tested https://github.com/apache/trafficserver/pull/8534 using NVDIMM, achieving a lag reduction.

[bryancall]

@gaowayne Yes, we don't have a way to tier the block storage for ATS. You can partition your cache based on hostnames if that help and store certain domains on the NVDIMM block devices. https://docs.trafficserver.apache.org/admin-guide/storage/index.en.html#partitioning-the-cache

[cukiernik]

@gaowayne .Can you verify and share a benchmark to https://github.com/cukiernik/trafficserver/tree/devdax?

[gaowayne]

@gaowayne .Can you verify and share a benchmark to https://github.com/cukiernik/trafficserver/tree/devdax? buddy, could you please share me quick starter benchmark steps? we can test this and share result here.

[cukiernik]

For monitoring IPM You can use: watch tools/ipm-performance Simple test:

seq 1 1000000 | while read N ; do
 echo "http://127.0.0.1:8080/${RANDOM}"
done | time xargs -P100 -n1 wget -O /dev/null --no-check-certificate 

[cukiernik]

Hi @gaowayne I completed a autoreconf and rules to initialize devices, but it still waiting for https://github.com/apache/trafficserver/pull/8534

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. Marking it stale to flag it for further consideration by the community.

[gaowayne]

@gaowayne .Can you verify and share a benchmark to https://github.com/cukiernik/trafficserver/tree/devdax?

sorry, we have give up Optane. I am so sorry for late response, I just read this. is it possible ATS do some tuning that leverage QLC SSD well. its seq write is 3.5GB/s random read is 6GB/s. if we do some improvement on the SW together. I can help do benchmark and we can release white paper together? we have slack to talk detail?

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. Marking it stale to flag it for further consideration by the community.