tera vs hbase

[caijieming-ng]

roadmap:

• [x] setup hbase 1.2.3 on hadoop 2.6
• [ ] hbase performance test on ycsb (headline 11.14)
• [ ] tera performance test on ycsb (headline 11.20)

计划：

• [x] 在hadoop 2.6上搭建hbase 1.2.3；
• [ ] 使用ycsb测试hbase性能；（dead line：11.14）
• [ ] 开发ycsb上对接tera的接口层；（dead line：11.20），并提pr到ycsb；

[caijieming-ng]

测试发现hbase随机写在抖动幅度比较大，每行1M，写带宽在100~400M之间抖动。 调研其compact策略：https://www.ngdata.com/visualizing-hbase-flushes-and-compactions/ 与rocksdb的Universal compaction策略一致 ref： https://github.com/facebook/rocksdb/wiki/Universal-Compaction

[caijieming-ng]

compact过程仿真代码：

#include <stdlib.h>
#include <iostream>
#include <vector>

int main(int ac, char* av[]) {
if (ac != 5) {
    std::cout << "Usage: " << av[0] << " flush_nr(10) flush_size(80M) min_compact(3) max_compact(10)\n";
    std::cout << "Usage Sample: " << av[0] << " 10 80 3 10\n";
    return 0;
}
int flush_nr = atoi(av[1]);
int flush_size = atoi(av[2]);
int min_compact = atoi(av[3]);
int max_compact = atoi(av[4]);
std::vector<int> storefiles;
for (int i = 0; i < flush_nr; i++) {
    // flush stage
    storefiles.push_back(flush_size);

    std::cout << "Store round[" << i << "]: [flush] ";
    for (int j = 0; j < storefiles.size(); j++) {
        std::cout << storefiles[j] << "    ";
    }
    std::cout << " ===> [compact] ";

    // check compact
    for (int j = 0; (storefiles.size() > min_compact) && (j < (int)storefiles.size() - min_compact); j++) {
        int current_size = storefiles[j];
        int sum = 0;
        for (int k = j + 1; k < (int)storefiles.size(); k++) {
            sum += storefiles[k];
        }
        if (current_size <= 1.2 * sum) {
            int nr_compact = 0;
            if (storefiles.size() - j - 1 > max_compact) {
                nr_compact = max_compact;
            } else if (storefiles.size() - j - 1 < min_compact) {
                nr_compact = 0;
            } else {
                nr_compact = storefiles.size() - j - 1;
            }
            // do compact
            if (nr_compact > 0) {
                int store_size = storefiles[j];
                for (int m = j + 1; m < nr_compact + j + 1; m++) {
                    store_size += storefiles[m];
                    storefiles[m] = storefiles[m + nr_compact];
                }
                storefiles[j] = store_size;

                int current_num = storefiles.size();
                storefiles.resize(current_num - nr_compact);
                break;
            }
        }
    }

    // stat current store file
    for (int j = 0; j < storefiles.size(); j++) {
        std::cout << storefiles[j] << "    ";
    }
    std::cout << std::endl;
}
return 0;
}

[caijieming-ng]

场景：每次flush 80MB的memtable，执行若干轮之后的结果： [xxx ~]$ ./hbase Usage: ./hbase flush_nr(10) flush_size(80M) min_compact(3) max_compact(10) Usage Sample: ./hbase 10 80 3 10 [xxx ~]$ ./hbase 10 80 3 10 Store round[0]: [flush] 80 ===> [compact] 80
Store round[1]: [flush] 80 80 ===> [compact] 80 80
Store round[2]: [flush] 80 80 80 ===> [compact] 80 80 80
Store round[3]: [flush] 80 80 80 80 ===> [compact] 320
Store round[4]: [flush] 320 80 ===> [compact] 320 80
Store round[5]: [flush] 320 80 80 ===> [compact] 320 80 80
Store round[6]: [flush] 320 80 80 80 ===> [compact] 320 80 80 80
Store round[7]: [flush] 320 80 80 80 80 ===> [compact] 640
Store round[8]: [flush] 640 80 ===> [compact] 640 80
Store round[9]: [flush] 640 80 80 ===> [compact] 640 80 80

[caijieming-ng]

可以分析一下该策略的写放大 :-)