VDS performance testing

I pick up the VDS hard - I thought about the performance comparison issue.

The purpose of this article is to try to find a criterion by which you can compare VDS from different providers and choose objectively the most successful price / quality combination. Perhaps the methods outlined in the article are not reliable, but as a starting point - quite enough.


To begin with, I see this question, if anything - correct me, because I am rather guided by common sense, because I have no experience with such tests.
')
On the one hand there is, for example, the benchmark “ubench”, which rather well, albeit surrogate, measures the power in parrots. It should be noted that in the case of VDS, all such testers are not entirely correct, but if there are sufficient statistics, the situation becomes quite close to real.

However, I thought to use another way. There is a CMS that generates some pages of the site for n seconds. I wrote a short script that requests 10 identical pages from different hosts, and then processes the result. Here is the most interesting thing - I personally am not quite sure of the correctness and adequacy of the methodology, therefore I will give the source codes of the scripts:

Script disp.php - tunnel for AJAX, I decided not to bathe with the cross-domain AJAX :)

<? print file_get_contents($_SERVER["QUERY_STRING"]) * 1000; ?>

Script exec.php - runs the main CMS script, emulating the call, measures the time of the call and displays it

<?

$time = microtime(1);
// HTTP
file_get_contents("http://{$_SERVER["HTTP_HOST"]}/");
print microtime(1) - $time;

?>

Actually, the most important script is located at test.osmio.ru/disp.html - all JavaScript code, source code is open.

Personally, I see several potentially slippery places:

1. If one script has not yet worked, and the new request has already arrived - the server starts to load harder, but the tick of synchronous requests will save it.
2. The script does not measure the total time of the request, although it might have followed - all servers respond with different delay ... But I deliberately did not include this in the calculation, since channels between servers can go as you please, including through Europe - and it happens that two neighboring DCs in Msk drive traffic roundabout ways.
3. HTTP request in exec.php - despite the fact that it goes to the same host

After receiving the required number of tests, I process them slightly, as far as the memories of the matane are enough, and draw them out.

Server configurations are as follows:

• vds-dbbb.1gb.ru - 1GB.ru, 500MHz, 384MB, OpenVZ, Gentoo
• cms.dis.dj - Infobox.ru, 333MHz, 512MB, HyperV, Debian5
• 89.188.123.24 - Ruweb.net, 2000MHz, 1536MB, VDSManager, FreeBSD6

I measured the following values:

Host	The average	Deviation	Min	Max	0	one	2	3	four	five	6	7	eight	9
vds-dbbb.1gb.ru	156	44 (28%)	128	282	136	138	131	282	137	134	134	133	144	128
cms.dis.dj	1294	60 (5%)	1234	1462	1275	1291	1279	1259	1234	1462	1282	1269	1262	1316
89.188.123.24	441	15 (3%)	417	460	451	420	452	456	445	460	417	445	420	443

I repeat once again that the methodology is not a fact that it is adequate to reality, but I will still draw some conclusions:

• 1GB of robust sausage, although the values ​​are mostly minimal
• Infobox shamelessly slows down - there is a weaker processor and a fundamentally different virtualization technology, but not as much
• Ruweb is very stable and between about 1GB and ISPServer

Ubench test results:

• Infobox
  Ubench CPU: 27692
  Ubench MEM: 19291
  Ubench AVG: 23491
• 1GB
  Ubench CPU: 55762
  Ubench MEM: 44521
  Ubench AVG: 50141
• Ruweb - according to the figures, the car is evil, but it must also be more powerful, the declared characteristics are a cut above the rest
  Ubench CPU: 600191
  Ubench MEM: 269780
  Ubench AVG: 434985

Conclusions - not such a weak VDS in the Infobox, even despite the lower frequency - it competes well with ISPServer. However, the generation time is much longer. I believe that this is due to the settings of the software. It should be noted here that everywhere except 1GB there is an ISPManager with Apache configured by default, PHP5, etc. It should also be noted that there is a definite correlation between the values ​​of ubench and what my tester measured.

UPD. ISPServer ended the test period, and since its result was not particularly good at a rather impressive price for hosting - I excluded it from the calculation.