dimakuv
commented
2 years ago This test was run on an ubuntu 20.04 azure DC-series VMs.
I'm assuming these are the machines that have only 128MB of EPC (Enclave Page Cache)? You can easily check it via the is-sgx-available utility that comes with Gramine.
The size of EPC is very important for performance. You can read through this docs, for example: https://gramine.readthedocs.io/en/latest/devel/performance.html#choice-of-sgx-machine
There are also several other subtle issues with SGX machines that can contribute to bad performance. For example, some older SGX machines do not have RDTSC instruction enabled inside SGX enclaves -- this also deteriorates enclave performance.
I am wondering if this is the most efficient way to run the redis server on gramine?
A typical way to test Redis is to run the client (redis_benchmark) on another machine, connected via 1Gb or 10Gb Ethernet link with your main SGX machine (where redis-server runs). This setup emulates the real-world environments where applications from the Internet/other machines on the cluster connect to the Redis server.
Because you run both the client and the server on the same machine, you don't account for the overhead/latency of the network link. This setup does not reflect how things work in reality, and thus shows unnecessarily bad performance.
Are there any other configurations I might benefit from tweaking, to make this benchmark test run faster?
redis-benchmark by default updates a single key-value pair (if I remember correctly). In reality, Redis servers hold thousands and millions of key-value pairs. Also, the typical access pattern for Redis is 80% reads and 20% writes (or 95% reads and 5% writes) -- it depends on what kind of workload you want to simulate with your tests. There are many resources on the Internet on how to do proper benchmarking of databases, including how to benchmark Redis. For example, the classic one: https://redis.io/topics/benchmarks
If you want to get the best performance of Redis in Gramine-SGX, probably the best configuration will be something like this:
- Redis server runs on a powerful ICX server (Icelake server), with GBs of EPC
- Redis client is connected to Redis server via a 1Gb Ethernet link
- Redis client updates 10,000 key-value pairs, with small-sized keys (10-60 bytes) and medium-sized values (1-4KB)
- Redis client performs 80% reads and 20% writes
The above configuration will probably give you something like 5% overhead of Gramine-SGX in comparison to native (this is just my projection, I have no real numbers to back up my hypothesis). But please note that my proposed setup may be not too realistic (especially the 1Gbps network link -- in an enterprise cluster, the network link would be at least 10Gbps I guess).
salsal97
Description of the problem
On comparing the performance of the gramine redis server with native using the redis-benchmark tool, gramine comes to be almost 2.75 times slower than running with its native counterpart. This test was run on an ubuntu 20.04 azure DC-series VMs.
Steps to reproduce
Followed quickstart instructions to install gramine using apt -
Expected results
Results are as expected; however, I am wondering if this is the most efficient way to run the redis server on gramine? Are there any other configurations I might benefit from tweaking, to make this benchmark test run faster? Interested in any input.
Actual results
NA