如何提高包裹在ThreadLocal中的SimpleDateFormat的性能?

这是在具有24个内核的RHEL上的Java 7(51)上,我们注意到随着线程池大小的增加,包裹在线程本地的java SimpleDateFormat的平均响应时间增加。这是意料之中的吗?或者,我只是在做一些愚蠢的事情?

enter image description here

测试程序

    public class DateFormatterLoadTest {
        private static final Logger LOG = Logger.getLogger(DateFormatterLoadTest .class);
        private final static int CONCURRENCY = 10;

        public static void main(String[] args) throws Exception {
            final AtomicLong total = new AtomicLong(0);
            ExecutorService es = Executors.newFixedThreadPool(CONCURRENCY);
            final CountDownLatch cdl = new CountDownLatch(CONCURRENCY);
            for (int i = 0; i < CONCURRENCY; i++) {
                es.execute(new Runnable() {
                    @Override
                    public void run() {
                        try {
                            int size = 65000;
                            Date d = new Date();

                            long time = System.currentTimeMillis();
                            for (int i = 0; i < size; i++) {
                                String sd = ISODateFormatter.convertDateToString(d);
                                assert (sd != null);
                            }
                            total.addAndGet((System.currentTimeMillis() - time));

                        } catch (Throwable t) {
                            t.printStackTrace();
                        } finally {
                            cdl.countDown();
                        }
                    }
                });
            }
            cdl.await();
            es.shutdown();
            LOG.info("TOTAL TIME:" + total.get());
            LOG.info("AVERAGE TIME:" + (total.get() / CONCURRENCY));
        }
    }

日期格式化程序类:

public class ISODateFormatter {
    private static final Logger LOG = Logger.getLogger(ISODateFormatter.class);

    private static ThreadLocal<DateFormat> dfWithTZ = new ThreadLocal<DateFormat>() {
        @Override
        public DateFormat get() {
            return super.get();
        }

        @Override
        protected DateFormat initialValue() {
            return new SimpleDateFormat("yyyy-MM-dd'T'HH:mm:ssZ",
                    Locale.ENGLISH);
        }

        @Override
        public void remove() {
            super.remove();
        }

        @Override
        public void set(DateFormat value) {
            super.set(value);
        }

    };

    public static String convertDateToString(Date date) {
        if (date == null) {
            return null;
        }
        try {
            return dfWithTZ.get().format(date);
        } catch (Exception e) {
            LOG.error("!!! Error parsing dateString: " + date, e);
            return null;
        }
    }
}

有人建议取出AtomicLong,所以只是想分享一下,它在增加平均时间方面没有任何作用:

##NOT USING ATOMIC LONG##
2014-02-28 11:03:52,790 [pool-1-thread-1] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:328
2014-02-28 11:03:52,868 [pool-1-thread-6] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:406
2014-02-28 11:03:52,821 [pool-1-thread-2] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:359
2014-02-28 11:03:52,821 [pool-1-thread-8] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:359
2014-02-28 11:03:52,868 [pool-1-thread-4] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:406
2014-02-28 11:03:52,915 [pool-1-thread-5] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:453
2014-02-28 11:03:52,930 [pool-1-thread-7] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:468
2014-02-28 11:03:52,930 [pool-1-thread-3] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:468
2014-02-28 11:03:52,930 [main] INFO  net.ahm.graph.DateFormatterLoadTest  - CONCURRENCY:8

##USING ATOMIC LONG##
2014-02-28 11:02:53,852 [main] INFO  net.ahm.graph.DateFormatterLoadTest  - TOTAL TIME:2726
2014-02-28 11:02:53,852 [main] INFO  net.ahm.graph.DateFormatterLoadTest  - CONCURRENCY:8
2014-02-28 11:02:53,852 [main] INFO  net.ahm.graph.DateFormatterLoadTest  - AVERAGE TIME:340

##NOT USING ATOMIC LONG##
2014-02-28 11:06:57,980 [pool-1-thread-3] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:312
2014-02-28 11:06:58,339 [pool-1-thread-8] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:671
2014-02-28 11:06:58,339 [pool-1-thread-4] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:671
2014-02-28 11:06:58,307 [pool-1-thread-7] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:639
2014-02-28 11:06:58,261 [pool-1-thread-6] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:593
2014-02-28 11:06:58,105 [pool-1-thread-15] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:437
2014-02-28 11:06:58,089 [pool-1-thread-13] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:421
2014-02-28 11:06:58,073 [pool-1-thread-1] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:405
2014-02-28 11:06:58,073 [pool-1-thread-12] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:405
2014-02-28 11:06:58,042 [pool-1-thread-14] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:374
2014-02-28 11:06:57,995 [pool-1-thread-2] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:327
2014-02-28 11:06:57,995 [pool-1-thread-16] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:327
2014-02-28 11:06:58,385 [pool-1-thread-10] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:717
2014-02-28 11:06:58,385 [pool-1-thread-11] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:717
2014-02-28 11:06:58,417 [pool-1-thread-9] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:749
2014-02-28 11:06:58,418 [pool-1-thread-5] INFO  net.ahm.graph.DateFormatterLoadTest  - THREAD TIME:750
2014-02-28 11:06:58,418 [main] INFO  net.ahm.graph.DateFormatterLoadTest  - CONCURRENCY:16

##USING ATOMIC LONG##
2014-02-28 11:07:57,510 [main] INFO  net.ahm.graph.DateFormatterLoadTest  - TOTAL TIME:9365
2014-02-28 11:07:57,510 [main] INFO  net.ahm.graph.DateFormatterLoadTest  - CONCURRENCY:16
2014-02-28 11:07:57,510 [main] INFO  net.ahm.graph.DateFormatterLoadTest  - AVERAGE TIME:585

答案 1

创建 SimpleDateFormat 的实例非常昂贵(本文介绍了一些分析/基准测试)。如果这是真的,与将日期解析为字符串相比,那么随着线程数的增加(以及因此增加SimpleDateFormat实例的数量,因为它们是线程局部的实例),您的平均时间将会增加。


答案 2

加快格式化速度的另一种方法是缓存格式化的结果。这考虑了这样一个事实,即通常没有那么多不同的日期需要格式化。如果拆分日期和时间的格式,它甚至是更好的缓存候选项。

这样做的缺点是,正常的Java缓存实现,如EHCache,会减慢速度,缓存访问只需要比格式化更长的时间。

还有另一个缓存实现,其访问时间与HashMap相当。在这种情况下,您可以获得不错的速度。在这里,您可以找到我的概念验证测试:https://github.com/headissue/cache2k-benchmark/blob/master/zoo/src/test/java/org/cache2k/benchmark/DateFormattingBenchmark.java

也许这可以成为你的方案中的解决方案。

免责声明:我正在处理缓存2k....


推荐