How to get access to Meta real traces

[wangyuyue]

Dear developers,
I want to test CacheLib using your real traces and analyze the distribution and changes of the traces.
But as I do aws s3 ls s3://cachelib-workload-sharing/pub/kvcache/202206/, it says Unable to locate credentials. You can configure credentials by running "aws configure"..

Does this mean the traces are not publically accessible yet? If not, how can I use it. And is there any other traces made public under the cachelib-workload-sharing directory? Thanks in advance.

[wangyuyue]

I think it's not any random aws configure that can access it.

[jaesoo-fb]

For QPS (Queries Per Second) control, we recommend to use the opRatePerSec (used by token bucket controller) which is set to 550K QPS by default.

[wangyuyue]

How about the OpDelayNs? Can I just make it zero? To be honest, I don't really understand the difference between OpDelayNs and opRatePerSec. It would be super helpful if you can briefly explain it. Thanks.

[jaesoo-fb]

Yep, just leave OpDelayNs to 0. Note that setting it along with the opRatePerSec will just make it hard to control the desired rate.

[wangyuyue]

And opRatePerSec controls the maximum rate, right? I mean if the CacheLib is unable to process so many queries in time, the input queries will be stalled instead of being discarded?

[jaesoo-fb]

Yes, right; opRatePerSec throttles the processing rate (as opposed to arrival rate) to be less than the max rate. If CacheLib is slow to process than the max rate, the processing of input queries will be stalled. There is no discard features.

See https://github.com/facebook/CacheLib/blob/main/cachelib/cachebench/runner/CacheStressor.h#L286 for reference; the throttleFn just adds some delays to slow down the processing.

[1a1a11a]

Thank you for open-sourcing the traces! Where can I find more information about the traces (e.g., the time length of the trace, what type of the service it is)? Thank you！

[jaesoo-fb]

Hi @1a1a11a

You can find some limited information from here, where we will keep appending info for any new traces later.

[1a1a11a]

Thank you! This is very helpful! One more question is the 930GB space dedicated to the workload, and is it still true that half of the SSD is used?

[wangyuyue]

Hi,
I find there is a configuration named cacheDir in the kvcache trace's configuration file. Could you give me some information about it? What effect does it have on the experiment? Thanks very much.

[wonglkd]

Probably this?

CacheLib/cachelib/cachebench/cache/Cache.h

Lines 108 to 109 in fbf277c

	// @param cacheDir optional directory for the cache to enable
	// persistence across restarts.

CacheLib/cachelib/cachebench/util/CacheConfig.h

Lines 68 to 70 in fbf277c

	// if set, we will persist the cache across cachebench runs. The directory
	// is used to store some metadata about the cache.
	std::string cacheDir{""};

[wangyuyue]

Thanks, I should have read the code more carefully :)