Introduction

How much faster is I/O from ramzswap compared to hard-disk? To determine this we check how long it takes for individual swap R/W operations to complete under heavy swapping conditions.

Details

Test config

• RAM: 2G
• CPU: Intel Core2 CPU 6400 @ 2.13GHz
• Swap partition (hard-disk): 4G
• Western Digital Raptor (WD1600ADFS)
• 10000 RPM
• 16MB Cache SATA 3.0Gb/s 3.5" Hard Drive
• ramzswap: 512MB (without backing swap) - swap usage during testing never exceeds 410MB.

Test details

• start swap systemtap probe script (see below).
• run scan benchmark to induce heavy swap I/O - it does scans of anonymous mapped memory, both cyclic and use once.
• ./scan 1800 10 (i.e. allocate 1800MB and scan over it 10 times).
• background tasks: window manager(kde4), music player(audacious), terminal(konsole).
• stop systemtap script once above scan command completes. On termination it shows swap I/O timing stats.

We use SystemTap to determine time taken by individual swap I/O. Following is the systemtap script used:

global time_read
global time_read_aggr

global time_write
global time_write_aggr

probe begin {
        print("swap probe begin\n")
}

probe end {
        print("swap probe end\n")
}

/* swap read */
probe kernel.function("swap_readpage") {
        time_read[tid()] = gettimeofday_ns();
}

probe kernel.function("end_swap_bio_read") {
        if (time_read[tid()]) {
                time_read_aggr <<< (gettimeofday_ns() - time_read[tid()])
                delete time_read[tid()]
        }
}

/* swap write */
probe kernel.function("swap_writepage") {
        time_write[tid()] = gettimeofday_ns();
}

probe kernel.function("end_swap_bio_write") {
        if (time_write[tid()]) {
                time_write_aggr <<< (gettimeofday_ns() - time_write[tid()])
                delete time_write[tid()]
        }
}

Output

For disk swap

time_read_aggr @count=0xc465 @min=0x1e1d @max=0x2778b69d5ef @sum=0x7b4718532b7 @avg=0xa0b13c9
time_write_aggr @count=0x1bb6 @min=0x1d0e @max=0x773d04af3e @sum=0x24b9c4f02a9 @avg=0x15347b3c

For ramzswap

time_read_aggr @count=0x412fba @min=0x59b @max=0x195473f @sum=0xc02d2b158 @avg=0x2f2b
time_write_aggr @count=0x414e35 @min=0x929 @max=0x15fcacd @sum=0x6fbdbb071 @avg=0x1b60

Results

Comparing average R/W times:

Notes

• It is surprising that ramzswap average read time > write time. LZO decompression is much faster than compression. Maybe there are some bottlenecks in ramzswap_read()?