NodeJS通过流复制文件非常慢

Question

我在VMWare下的SSD上用Node复制文件,但性能非常低.我测量实际速度的基准如下:

$ hdparm -tT /dev/sda

/dev/sda:
 Timing cached reads:   12004 MB in  1.99 seconds = 6025.64 MB/sec
 Timing buffered disk reads: 1370 MB in  3.00 seconds = 456.29 MB/sec

但是,以下复制文件的节点代码非常慢,因此后续运行不会使其更快:

var fs  = require("fs");
fs.createReadStream("bigfile").pipe(fs.createWriteStream("tempbigfile"));

运行如下:

$ seq 1 10000000 > bigfile
$ ll bigfile -h
-rw-rw-r-- 1 mustafa mustafa 848M Jun  3 03:30 bigfile
$ time node test.js 

real    0m4.973s
user    0m2.621s
sys     0m7.236s
$ time node test.js 

real    0m5.370s
user    0m2.496s
sys     0m7.190s

这里有什么问题,如何加快速度？我相信只需调整缓冲区大小,我就可以在C中更快地编写它.令我困惑的是,当我编写简单的几乎pv等效程序时,将stdin管道输出到stdout,如下所示,它非常快.

process.stdin.pipe(process.stdout);

运行如下:

$ dd if=/dev/zero bs=8M count=128 | pv | dd of=/dev/null
128+0 records in 174MB/s] [        <=>                                                                                ]
128+0 records out
1073741824 bytes (1.1 GB) copied, 5.78077 s, 186 MB/s
   1GB 0:00:05 [ 177MB/s] [          <=>                                                                              ]
2097152+0 records in
2097152+0 records out
1073741824 bytes (1.1 GB) copied, 5.78131 s, 186 MB/s
$ dd if=/dev/zero bs=8M count=128 |  dd of=/dev/null
128+0 records in
128+0 records out
1073741824 bytes (1.1 GB) copied, 5.57005 s, 193 MB/s
2097152+0 records in
2097152+0 records out
1073741824 bytes (1.1 GB) copied, 5.5704 s, 193 MB/s
$ dd if=/dev/zero bs=8M count=128 | node test.js | dd of=/dev/null
128+0 records in
128+0 records out
1073741824 bytes (1.1 GB) copied, 4.61734 s, 233 MB/s
2097152+0 records in
2097152+0 records out
1073741824 bytes (1.1 GB) copied, 4.62766 s, 232 MB/s
$ dd if=/dev/zero bs=8M count=128 | node test.js | dd of=/dev/null
128+0 records in
128+0 records out
1073741824 bytes (1.1 GB) copied, 4.22107 s, 254 MB/s
2097152+0 records in
2097152+0 records out
1073741824 bytes (1.1 GB) copied, 4.23231 s, 254 MB/s
$ dd if=/dev/zero bs=8M count=128 | dd of=/dev/null
128+0 records in
128+0 records out
1073741824 bytes (1.1 GB) copied, 5.70124 s, 188 MB/s
2097152+0 records in
2097152+0 records out
1073741824 bytes (1.1 GB) copied, 5.70144 s, 188 MB/s
$ dd if=/dev/zero bs=8M count=128 | node test.js | dd of=/dev/null
128+0 records in
128+0 records out
1073741824 bytes (1.1 GB) copied, 4.51055 s, 238 MB/s
2097152+0 records in
2097152+0 records out
1073741824 bytes (1.1 GB) copied, 4.52087 s, 238 MB/s

Answer 1

我不知道你的问题的答案,但也许这有助于你调查问题.

在关于流的Node.js文档中,在Streams Under the Hood:Buffering下它说:

Writable和Readable流都将分别在名为_writableState.buffer或_readableState.buffer的内部对象上缓冲数据.

可能缓冲的数据量取决于传递给构造函数的highWaterMark选项.

[...]

流的目的,特别是使用pipe()方法,是将数据缓冲限制在可接受的水平,以便不同速度的源和目标不会压倒可用内存.

因此,您可以使用缓冲区大小来提高速度:

var fs = require('fs');
var path = require('path');
var from = path.normalize(process.argv[2]);
var to = path.normalize(process.argv[3]);

var readOpts = {highWaterMark: Math.pow(2,16)};  // 65536
var writeOpts = {highWaterMark: Math.pow(2,16)}; // 65536  

var source = fs.createReadStream(from, readOpts);
var destiny = fs.createWriteStream(to, writeOpts)

source.pipe(destiny);