A sketch is a probabilistic data structure used to record frequencies of items in a multi-set. Sketches are widely used in various fields, especially those that involve processing and storing data streams. In streaming applications with high data rates, a sketch "fills up" very quickly. Thus, its contents are periodically transferred to a remote collector, which is responsible for answering queries. In this paper, we propose a new sketch, called Slim-Fat (SF) sketch, which has a significantly higher accuracy compared to prior art and at the same time achieves comparable speed as the best prior sketch. The key idea behind our proposed SF-sketch is to maintain two separate sketches: a large sketch called Fat-subsketch and a small sketch called Slim-subsketch. The Fat-subsketch is used to perform updates, and periodically produce the Slim-subsketch, which is periodically transferred to the remote collector for answering queries quickly and accurately. For the final version of our SF-sketch, we derived the error bound and an accurate correct rate formula verified by experiments. We implemented and extensively evaluated SF-sketch along with several prior sketches and compared them side by side. Our experimental results show that SF-sketch outperforms the most widely used CM-sketch by up to 33.1 times in terms of accuracy.
We implement SF-sketch and 4 well known sketches (CMCU sketch, A sketch, CM sketch and C sketch) which are used for comparing with our proposed SF sketch. You can build thoses sketches by
cd [sketch folder]; ./cmkae .; make
There is an example in main.c, which shows the basic usage of the corresponding sketch.
Type [sketch folder]/bin/[sketch] -h to show more information.
NOTICE: CMake 3.1 or higher is required
$ cd ./bench_sketch/; make;
$ sh ./experiment_on_timming.sh
$ sh ./experiment_on_aging.sh
To test the performance of sketches in different scenarios, we harness YCSB to generate two kinds of workloads: uniform and skewed (zipfian). We also use memcached and the generated workloads to record the real frequency of each item as benchmarks. After that, we feed the generated workloads to our proposed SF-sketch to record the estimation of itme frequency using the SF-sketch. Then, we calculate the average relative error and empirical cumulative distribution function (CDF) of relative error using benchmarks and estimations. To compare the accuracy with other well known sketches (such as Count-min (CM) sketch, conservative update (CU) sketch, etc.), we can use the same method to get the average relative error and CDF.