Performance when changing the length of a keys value. (It gets really bad.)

[Alex Stapleton]

On 13 Jul 2005, at 17:27, Alex Stapleton wrote:

> I am running under Debian 3.0 with Linux 2.4.23, Quad 2.4 GHz P4  
> Xeon. 2GB RAM. My memcached is running with these options
>
> /usr/bin/memcached -m 128 -p 11211 -u root -c 1024
>
> Performance is generally rather good. I can push 6000-7000 requests/ 
> s doing over 100MB/s of data in throughput. However! In my test  
> code, if I change the data block size from 8192 bytes, to 1024  
> bytes, without restarting the server, Performance suddenly becomes  
> hideously slow. So slow in fact, the tests which where taking only  
> 2-3 seconds to complete, now take over 10 times longer, reducing my  
> requests a second from several thousand, to below 10!
>
> Does anyone have any suggestions as to why this might be? I am  
> writing approximately 80 MB of data to the server in my tests,  
> using the same keys between tests, so they should just get  
> overwritten, right? I expect it is due to memcached suddenly  
> deciding to re-malloc or free memory when I reduce the data size.
>
> Here is my test script. It is written in Python using the danga  
> memcached module. I hope nothing mangles the formatting.
>
> import time
> import memcache
>
> mc = memcache.Client(['127.0.0.1:11211'], debug=0)
>
> iters = 100
> data_len = 4098
>
> data = "B" * data_len
>
> s = time.time()
>
> for i in xrange(1,iters):
>         mc.set(i, data)
>
> e = time.time()
> took = (e-s)
>
> print "Took %s seconds" % (took)
> print "Wrote at least %s MB of data" % ((data_len*iters)/1024/1024)
> print "%s t/s" % (iters/took)
> print "%s s/t" % (took/iters)
> print "%s MB/s" % ((data_len*iters/took)*2/1024/1024)
>

Hmmm, I thought -k fixed this, but it seems to just make the issue  
less obvious. Now if I switch between extremes (31Kb blocks (32KB  
ones are VERY slow) to 1 byte blocks) the performance become  
increadibly bad again! With -k enabled I don't see why it should be  
doing this.