Help on server performance

[paul c]

We currently run two servers (UNIX Intel P4 2.53 GHz) with a load balance addition. Whilst we’ve been told this is more than suitable for our needs, we are finding this is not the case. We use the servers for simple swf file downloads and nothing else, a typical file will vary in size from 150k to 750k. Speed of initial file download start is business critical. When we are running at 10Mb/s across the two servers everything is fine – a file will start to download in 1 second on broadband and 5 seconds on dial-up. But when we have a heavy day this will peak at 60-80 Mb/s across the two servers and file call up slows down (5 & 15 seconds respectfully) and we get interrupted file downloads. We are told the servers could cope with 120 Mb/s combined, but at half this we find the servers are struggling. Have we got the right set up? Do we need more powerful servers? Do we need more (loathed to do this as our business could require endless servers as it grows)? Are we right to use dedicated servers – should we be on a distributed server network?

[dynamicnet]

Greetings Paul:

How much memory is in each server?

What does the server load look like at peak?

Do you have 100 MB per second capable network cards in the servers?

Are the servers connected to a 100 MB per second port?

Are you using hardware load balancing?

Load balancing offers various settings as to know when to switch between servers; what setting are you using?

What carriers are you using for connectivity (bandwidth)?

Thank you.

[paul c]

The servers are: BioStar MB, Intel P4 2.53 GHz, 1 GB DDR RAM, 40 GB IDE, FreeBSD OS. We use Two load balanced ports.
Carriers are UUNet, Qwest, Level 3, Global Crossing and MCI.
I hope we have 100 Mb/s cards and ports - I assumed so but will ask.

[profitability]

A few notes:

1. Are you sure that it's the servers not coping with the load? Are there any firewalls/proxies/etc. upstream from the servers that might be choking on the heavier load?

2. Not to start an IDE vs. SCSI religious war here, but for high-performance applications such as this, investing in SCSI technology might not be a bad idea, particularly if your application is read/write intensive.

3. When you start getting up around 80Mb/s you might want to start considering moving to Gigabit Ethernet -- you might simply be maxing out the capabilities of your NIC cards.

Andrew

[paul c]

Thanks -
1. No I'm not sure it's the servers - but we keep raising the issuew with our hoisting company and they've not suggested anything about firewalls
2 IDE vs SCSI - neither mean anything to me I'm afraid - I'm the business driver rather than tech ewxpert we've been relying on hosting company for expertese. Could you explain?
3. The NIC card does make sense I'll check.
Paul

[dynamicnet]

Greetings Paul:

I'm more of a software engineer than a hardware tech; for me, SCISI hard drives are just plain faster under most (if not all) situations.

IDE hard drives typically max out at 7,500 RPM; while SCSI hard drives start out at 10,000 RPM; today it is easy to find 15,000 RPM SCSI drives.

Something you may already be checking on as well, but in the 1st post I asked what method is the load balancing using to determine which server gets hit.

I could be that under certain conditions, it is picking a server already maxed out.

Thank you.

[mrbeans]

A few notes:
3. When you start getting up around 80Mb/s you might want to start considering moving to Gigabit Ethernet -- you might simply be maxing out the capabilities of your NIC cards.


I second with statement. With 100Mb/s ethernet, your realistic upper bound on throughput for many simultaneous connections is about 80Mb/s with the overhead etc... I would bet that if you are using a hardware load balancer, the outside interface which actually connects to the internet is being saturated. I would find out what speed of interfaces you have in ALL of your equipment from the servers until it hits a backbone, and how fast your uplink to that backbone network is.

[paul c]

Hi Dynamicnet
Thanks for your response. RPM? Excuse my ignorance on a fast learning curve.
We have a hardware load balancer, and I suspected we had a bias to one server but not sure. We have two stats sources (i) MRTG that implied both servers were taking roughly the same level of hits (ii) Urchin campaign stats - this implied a heavy (very heavy bias) to one server (90%). When I pointed this out to the company that managers the servers they said MRTG was more accurate and blamed Urchin for loosing one of the servers stats. Paul

[paul c]

[QUOTE=mrbeans]I second with statement. With 100Mb/s ethernet, your realistic upper bound on throughput for many simultaneous connections is about 80Mb/s with the overhead etc... I would bet that if you are using a hardware load balancer, the outside interface which actually connects to the internet is being saturated.

In practise each individual server (according to MRTG stats) was only reaching 25 Mb/s second - well within their supposed capabilities - and the hosting company is suggesting we simply ad another server. But we were experiencing considerable slow down in response times when existing set up should be ok so am loathed to sign up to more. And yes we use a hardware load balancer.
Paul

[paul c]

We have a gig-fiber uplink to a hardware loadbalancing with Foundry ServerIrons - ours are the only servers connected to the load bancer. I'm told a dns round Robin would slow things down even more. Hosting company suggests move to SCSI is best bet for improving delivery - any thoughts

[dynamicnet]

Hi Paul:

RPM (like in a car so to speak) == revolutions per minute. The higher the number, the faster the drive.

I would check if the hardware load balancer is doing more than here, there, here, there, here, there, here there transfers. I.e. just switching back and forth between the two.

From our limited (I'm more of a software / operating system person) experience with Alteon load balancing hardware, you could set up various means to determine which server to go to from server load to number of processes, and more.

Thank you.

[profitability]

One other thought worth looking into...you don't mention how the files are being transfered -- http, ftp, something else? In any case, you might want to make sure that whatever daemon you're using to serve the files up is configured to be able to accept enough concurrent sessions to handle the load you're throwing at it. If you're running up against a session limit on your server you could see some of the symptoms you're describing.

Andrew