Amazon EC2 very slow?

Portal Home > Knowledgebase > Articles Database > Amazon EC2 very slow?

Posted by michael79, 05-26-2011, 08:42 AM
Hi, I have just setup a "Amazon Hi-CPU 8 Cores"-instance, running Linux 64-Bit. However, my webapp is much slower then on my dedicated server. Now I have tested "Unixbench" and it seems to confirm the rather poor performance. The Overall-Score for all 8 CPUs is 1958.3, for one CPU 574.4. Compared with my old dedicated server this is very slow (Intel Core i7-920 Quad-Core/ SATA-II HDD has a score of about 10000). However, the Amazon-Instance is already a high-end-version. Is Amazon-EC2 really so slow or am i Doing something wrong? Here are the Unixbench-details for EC2: Benchmark Run: Thu May 26 2011 11:24:34 - 11:52:47 8 CPUs in system; running 1 parallel copy of tests Dhrystone 2 using register variables 17969091.5 lps (10.0 s, 7 samples) Double-Precision Whetstone 2786.0 MWIPS (10.0 s, 7 samples) Execl Throughput 1462.3 lps (30.0 s, 2 samples) File Copy 1024 bufsize 2000 maxblocks 252737.9 KBps (30.0 s, 2 samples) File Copy 256 bufsize 500 maxblocks 68924.9 KBps (30.0 s, 2 samples) File Copy 4096 bufsize 8000 maxblocks 647660.0 KBps (30.0 s, 2 samples) Pipe Throughput 392265.1 lps (10.0 s, 7 samples) Pipe-based Context Switching 99087.0 lps (10.0 s, 7 samples) Process Creation 3726.4 lps (30.0 s, 2 samples) Shell Scripts (1 concurrent) 4164.3 lpm (60.0 s, 2 samples) Shell Scripts (8 concurrent) 1516.6 lpm (60.0 s, 2 samples) System Call Overhead 428797.3 lps (10.0 s, 7 samples) System Benchmarks Index Values BASELINE RESULT INDEX Dhrystone 2 using register variables 116700.0 17969091.5 1539.8 Double-Precision Whetstone 55.0 2786.0 506.5 Execl Throughput 43.0 1462.3 340.1 File Copy 1024 bufsize 2000 maxblocks 3960.0 252737.9 638.2 File Copy 256 bufsize 500 maxblocks 1655.0 68924.9 416.5 File Copy 4096 bufsize 8000 maxblocks 5800.0 647660.0 1116.7 Pipe Throughput 12440.0 392265.1 315.3 Pipe-based Context Switching 4000.0 99087.0 247.7 Process Creation 126.0 3726.4 295.7 Shell Scripts (1 concurrent) 42.4 4164.3 982.1 Shell Scripts (8 concurrent) 6.0 1516.6 2527.6 System Call Overhead 15000.0 428797.3 285.9 ======== System Benchmarks Index Score 574.4 ------------------------------------------------------------------------ Benchmark Run: Thu May 26 2011 11:52:47 - 12:21:43 8 CPUs in system; running 8 parallel copies of tests Dhrystone 2 using register variables 137758198.6 lps (10.0 s, 7 samples) Double-Precision Whetstone 22155.8 MWIPS (10.0 s, 7 samples) Execl Throughput 7674.6 lps (29.9 s, 2 samples) File Copy 1024 bufsize 2000 maxblocks 253551.7 KBps (30.0 s, 2 samples) File Copy 256 bufsize 500 maxblocks 68429.3 KBps (30.0 s, 2 samples) File Copy 4096 bufsize 8000 maxblocks 641536.4 KBps (30.0 s, 2 samples) Pipe Throughput 2924895.1 lps (10.0 s, 7 samples) Pipe-based Context Switching 719235.6 lps (10.0 s, 7 samples) Process Creation 20062.2 lps (30.0 s, 2 samples) Shell Scripts (1 concurrent) 16201.5 lpm (60.0 s, 2 samples) Shell Scripts (8 concurrent) 2377.8 lpm (60.1 s, 2 samples) System Call Overhead 1883265.0 lps (10.0 s, 7 samples) System Benchmarks Index Values BASELINE RESULT INDEX Dhrystone 2 using register variables 116700.0 137758198.6 11804.5 Double-Precision Whetstone 55.0 22155.8 4028.3 Execl Throughput 43.0 7674.6 1784.8 File Copy 1024 bufsize 2000 maxblocks 3960.0 253551.7 640.3 File Copy 256 bufsize 500 maxblocks 1655.0 68429.3 413.5 File Copy 4096 bufsize 8000 maxblocks 5800.0 641536.4 1106.1 Pipe Throughput 12440.0 2924895.1 2351.2 Pipe-based Context Switching 4000.0 719235.6 1798.1 Process Creation 126.0 20062.2 1592.2 Shell Scripts (1 concurrent) 42.4 16201.5 3821.1 Shell Scripts (8 concurrent) 6.0 2377.8 3963.0 System Call Overhead 15000.0 1883265.0 1255.5 ======== System Benchmarks Index Score 1958.3
Posted by dazmanultra, 05-26-2011, 08:43 AM
A single Amazon instance can be sluggish compared to real physical hardware - but then Amazon instances are really designed to be used for horizontally scaling an application (adding more instances, and distributing the load using clustering or load balancing) rather than vertical scaling (using one server, but making it faster).
Posted by Orien, 05-26-2011, 01:38 PM
It looks like your Unixbench test is also running the disk tests and that area isn't something EC2 will excel in.
Posted by tchen, 05-26-2011, 03:55 PM
Hey, looks like they've actually gotten better From mid 2010 (courtesy of cloudharmony): http://benchmarks.cloudharmony.com/u...ge-us-west.txt You probably ended up on one of the X5550's or better, but I don't think you'll get close to the 10000 unixbench score on any cloud provider just because that test behaves erratically inside a vm. It's one reason why cloud harmony runs a series of app tests instead of just the one. That said, you will see perceptible decrease in performance on heavily loaded vm node just because of cache load issues if you're sharing part of those 8 cores.
Posted by Marco van den Akker, 05-26-2011, 04:01 PM
I assume Lady GaGa and her release.
Posted by streaky81, 05-26-2011, 07:47 PM
I was ranting at sencha the other day because their site was slow as hell - turned out to be running on EC2 which said it all to me
Posted by seg fault, 06-03-2011, 10:57 AM
I have also been underwhelmed with EC2's performance. My dedicated server was idle 90% of the time, but that 10% when it was busy was the most important. EC2 just doesn't burst very well at all, if anything, they appear to be heavily oversubscribed
Posted by tchen, 06-03-2011, 11:15 AM
... probably because EC2 doesn't have burst (aside from micro instances).
Posted by schooglepets, 06-03-2011, 11:22 AM
Who would be a good alternative to EC2 then? I've had trouble finding a provider with as much flexibility as EC2.
Posted by tchen, 06-03-2011, 12:07 PM
It depends on what you're looking for. If its horizontal scaling for non-standard apps, then EC2 does provide the flexibility that's hard to beat. If your requirements skew towards the dedicated-server style setup, then EC2 is so wrong for you. If you're looking for a single dedicated server performance with basic failover, then you should just look at providers that use AppLogic. You'd want to make sure it isn't a shared-node system (obviously). If you're building a small cluster range with reasonable forecasts of load, then you can either build it yourself out of dedi/vps boxes or stick with AppLogic. EC2 can do these things, but it really isn't the most optimal for known quantities or small scales.
Posted by nancysmith, 06-18-2011, 07:17 AM
i think that EC2 will be best option for you.
Posted by FusionNET, 06-18-2011, 04:04 PM
I would check out CleverKite's new clouds, I hear there quite amazing and I had my eye out on it and am currently checking it out. I'll update you once I get mine and try it out, but I've heard nothing but crazy insane positive feed back about the disk speed, support, and the overall service.
Posted by boskone, 06-18-2011, 04:08 PM
EC2 is one of the worst performing clouds on the market at the moment...
Posted by schooglepets, 06-18-2011, 04:16 PM
Are they even out yet? I e-mailed them a week or so ago and they said they are still in beta testing.
Posted by HostHatch_AR, 06-18-2011, 05:42 PM
Same with me, Emailed them and they said "At the end of the month", Would love to try one though
Posted by Orien, 06-19-2011, 09:02 PM
Indeed, we planned to have everything released by now, but ran into some software compatibility issues that we're working on resolving right now. Appreciate the interest!
Posted by bergholt, 06-20-2011, 03:22 AM
Each vCPU on Amazon's EC2 platform gives you around processing power equivalent of 400 passmarks. Having access to 8 vCPU's is around 3200 passmarks. In comparison a i7-920 scores around 5.500. So yes, in terms of raw cpu power, your dedicated server setup was more powerful.
Posted by dazmanultra, 06-20-2011, 03:57 AM
I think you should read about parallel computing, and specifically Amdahl's law. Hosting is a very good example of a repetitive action that whilst each individual action might not be able to be sped up, you can serve many requests concurrently. Adding more servers allows you to serve more requests concurrently.

Add to Favourites Print this Article