Redis benchmark: multithread support and cluster support

[artix75]

This PR implements support for threads and cluster into redis-benchmark.

Multi-Thread support

By default, redis-benchmark will be launched in single-thread mode as usual.
In order to launch it in multi-thread mode, you can use the --threads option, ie:

redis-benchmark --threads 6

The command above will launch redis-benchmark with six threads.

Cluster Support

Redis benchmark can now work with clusters. This can be done by using the --cluster option.
The redis instance which redis-benchmark will connect to will be used as the cluster entry point.

Examples:

redis-benchmark --cluster -p 7000

The line above will launch redis-benchmark in cluster mode and will use the instance with port 7000 as the cluster entry point.

When launched in cluster mode, redis-benchmark will implicitly start in multi-threading mode too.
By default, it will use one thread for each master node found in the cluster.
Anyway, you're free to change this default configuration by explicitly pass the --threads option, ie:

redis-benchmark -p 7000 --threads 6 --cluster # If the cluster has three masters, use two threads per master node

When launched in cluster mode, you can use the special {tag} placeholder inside the key names in order to let redis-benchmark use slot hash tags, in a similar fashion of the __rand_int__ placeholder.
In this case, redis-benchmark will automatically replace the {tag} placeholder with the proper slot hash tag depending on which master node is querying.

Example:

redis-benchmark -p 7000 --cluster get mykey:{tag}

Default redis-benchmark's tests already have the {tag} placeholder in their key name(s).

You're free to mix the {tag} placeholder and the __rand_int__ placeholder in order to be sure that every thread uses random and different keys (remember to use the -r option to enable the key name randomization).

Example:

redis-benchmark -r -p 7000 --cluster get mykey:{tag}:__rand_int__

When working in cluster mode, redis-benchmark will always fetch the cluster configuration before starting the test(s). Anyway, if some slot is moved during the tests (ie. because of a resharding), after receiving an "ASK/MOVED" reply error, redis-benchmark will fetch the cluster configuration again in order to update it.
The configuration update will be atomically handled by threads.

[antirez]

Aweomse! Thanks Fabio, this feature will be super useful in the future, finally folks can see the numbers Redis Cluster can provide :-D

[antirez]

@artix75 P.S. Given that the PR description is so well put, maybe you should PR that into the redis-doc repository in the "benchmarking" documentation.

[artix75]

@artix75 P.S. Given that the PR description is so well put, maybe you should PR that into the redis-doc repository in the "benchmarking" documentation.

Yeah, i agree

[wmh]

-a <password> is no longer available with --cluster in this patch

[artix75]

-a <password> is no longer available with --cluster in this patch

@wmh See PR #6665

[wmh]

In cluster mode, it will stop after the test of LRANGE_600 with an error message:

LRANGE_600 (first 600 elements): 33156.50 requests per second
Error: Server closed the connection

It should following a MSET test result like this in non-cluster mode:

LRANGE_600 (first 600 elements): 110864.74 requests per second
MSET (10 keys): 108577.63 requests per second

I guess MSET is not support in cluster mode, so it would be better if we can just skip the test.