New Server Hardware and price quotes, Funding the server

[Angular]

Lets figure out what hardware is needed to run the 'new' GIMPS server well. Of course there are many different options and configurations. A separate web server and database server would be ideal, but the cost may be prohibitive.

I think there is little question that we should go with a new server with a good onsite service plan. GIMPS is too large to have much down time! The hardware should be as redundant as possible. Would you agree that a RAID5 (or RAID10,50,...) and Redundant power supplies are essential features?

This is the current server:

Originally Posted by Prime95

a dual 800Mhz PIII with 36GB fast RAID 5 disks (18GB usable) and 256MB, on a T1 or DS3 (not sure which) with Win2000.

1) Rackmount case: Can we settle on a 2U-4U rackmount box? Are there additional hosting charges for a 4U box? If so, please specify.

2) CPU: There are two main options for the CPUs: Dual Xeon or Dual Athlon MP [ or if cost is a concern Dual PIII, now up to 1.4GHz; but note that the PIII 1.4 ~ P4 2.4 cost].
What are the benchmarks for the Athlons and Xeons in mySQL? (or what DB are we planning on using?)

Originally Posted by Prime95

The principle CPU requirement is to drive the database. The web server runs on the same box for little overhead, but the SQL reports often eat almost 100% CPU for minutes. ... Transactions average 2 to 3/sec for nominal (around the clock average)client load, with spikes of up to several dozen per second.

George, what is the average processing time for each transaction?

3) Memory: I think the current 256MB of RAM in the current server is probably a bottleneck. It is very important to have sufficient RAM to completely cache the DB,

Originally Posted by Prime95

RAM makes a big performance difference since SQL Server caches pretty much the entire set of tables and only the write-thrus do I/O (thus reports are mostly CPU/bus bound)

How large is the DB and how much RAM would be needed to completely cache it? (room for future growth?)
I think 512Mb is the minimum we should look at. 1GB would be preferred if we have the financial resources.

4) Storage: How critical is the speed of our storage subsystem? I would say a 3 x 36GB configured in RAID 5 is the minimum option. Better would be with an additional 2x18GB RAID 1 for the OS. 15k or 10k HDs? A caching RAID controller would also be helpful.
A Hot spare for the RAID 5 would limit down time. How long would it take to get personal to the server to replace the HD? If we loose one drive from the RAID 5 array we would not loose data, but until the drive is replaced the array would be several times slower.

5) Service: A good service agreement is very important. Same Day or Next day service? Dell for example, offers a 3 year Same Day 4Hr response with parts/labor or Next Business day. Then we often have the option of business hrs response or 24/7 response.

I think that covers most of the hardware choices. Are there any other important hardware considerations? Adding an additional NIC for ~$60 would also help the redundancy of the system.

Now we need to discuss these requirements, gets server quotes, and find funding to purchase the server.

I am not saying we need to purchase the server now, but lets just get the details ready. Then when we need the server all we need to do is purchase it and set it up.

[ET_]

Originally Posted by Angular

2) CPU: There are two main options for the CPUs: Dual Xeon or Dual Athlon MP [ or if cost is a concern Dual PIII, now up to 1.4GHz; but note that the PIII 1.4 ~ P4 2.4 cost].
What are the benchmarks for the Athlons and Xeons in mySQL? (or what DB are we planning on using?)

3) Memory: I think the current 256MB of RAM in the current server is probably a bottleneck. It is very important to have sufficient RAM to completely cache the DB.

How large is the DB and how much RAM would be needed to completely cache it? (room for future growth?)
I think 512Mb is the minimum we should look at. 1GB would be preferred if we have the financial resources.

2)
I'd prefer Dual Athlon. They perform better than Xeon and PIII in the average work, unless you hand-code each application.

3)
If you use a transactional database, 1GB RAM would be better to cache all the tablespace. If you plan 512, you have to redesign the database control files when upgrading (Oracle docet).

Luigi 8)

[adpowers]

In terms of processors, I think we should go with dual Athlon MPs. It is has a much better price/performance ratio (last time I checked) than versus Xeons. If you would like to see the performance, you can look at this article:
http://www.anandtech.com/IT/showdoc.html?i=1641
If we were going to have this professional hosted (in a colo facility) we should probably go with a 1U case. This would save money because they do charge by the U. Anandtech also has a comparison of a bunch of 1U dual Athlon MP systems here:
http://www.anandtech.com/IT/showdoc.html?i=1695

In terms of hosting, it would be very expensive to send it to a colo. If we could find someone at a university that would be willing to host it, that would be ideal. Universities have lots of bandwidth and since this is a mathematical project, they should find it a worthy use of bandwidth. Also, if we were hosting in a college, we wouldn't neccessarily need a 1U case, because we wouldn't get charged by the U (hopefully, we wouldn't get charged at all).

For storage, I think SCSI RAID is overkill. Instead I think it should be IDE RAID. This is because SCSI discs are a lot more money than IDE disks and SCSI RAID controllers are really, really expensive (much more so than IDE RAID). Instead, I suggest we go for 1 gig of RAM. Since this would allow the server to cache most, if not all, of the database and since most DB requests are probably read requests (for generating the stats), you wouldn't need to use the hard drives very much.

If we don't have to put it in a 1U case, then we could probably build it cheaper than buying. You mention Dell, but we couldn't use Dell if we went with the Athlon MPs because Dell doesn't do AMD.

In terms of redundant NICs, a few of the Athlon MP motherboards have dual NICs built in. However, if you want ultimate redundancy, then I think proxy servers would be the best idea. Sure, the stats website would go down, but people would still be able to reserve and return exponents, which is the most important.

These are my thoughts,
Andrew

[Angular]

Originally Posted by adpowers

If we were going to have this professional hosted (in a colo facility) we should probably go with a 1U case. This would save money because they do charge by the U.

That is a valid point. The storage subsystem must fit in the 1U case. How many hard drives can a 1U case contain?

For storage, I think SCSI RAID is overkill. Instead I think it should be IDE RAID. This is because SCSI discs are a lot more money than IDE disks and SCSI RAID controllers are really, really expensive (much more so than IDE RAID).

How do you plain to do IDE RAID 5? They make controllers but are they integrated into any Dual Athlon MB that fit in a 1/2 U case? Or will the IDE Raid PCI card fit into a xU rack case?

RAID 5 is essential. We can not afford to loose data from a HD crash. Recall that most IDE drives only have a 1 year warranty now. [some high end drives have the 3 year warranty] A 5 year warranty is standard on SCSI drives. A 36GB 10,000 rpm SCSI drive is only $160. SCSI has 15k rmp drives. IDE still does not have any 10k rpm drives.

I agree 1GB is Ram is a wise choice.

If we don't have to put it in a 1U case, then we could probably build it cheaper than buying. You mention Dell, but we couldn't use Dell if we went with the Athlon MPs because Dell doesn't do AMD.

Yes. I just mentioned there service options.

Since we need service, what are good but cheap Athlon MP rack mount vendors?

[Angular]

Originally Posted by ET_

2)
I'd prefer Dual Athlon. They perform better than Xeon and PIII in the average work, unless you hand-code each application.

3)
If you use a transactional database, 1GB RAM would be better to cache all the tablespace. If you plan 512, you have to redesign the database control files when upgrading (Oracle docet).

Luigi 8)

I don't mind Athlons, I just want to see some mySQL benchmarks on Xeons with dual channel DDR / RDRAM and Athlon MPs.

Do you need 1GB of _available_ RAM to cache all of the tablespace?

[Angular]

Originally Posted by adpowers

In terms of processors, I think we should go with dual Athlon MPs. It is has a much better price/performance ratio (last time I checked) than versus Xeons. If you would like to see the performance, you can look at this article:
http://www.anandtech.com/IT/showdoc.html?i=1641

Yes I recall that article. I am a bit lerry of their homegrown DB tests. Woud our DB transactions be similar to a Forum?? What DB system are they using? To confirm their results if they use the DB we want to use (mySQL??) are their any other Athon vs. Xeon benchmarks?

In terms of hosting, it would be very expensive to send it to a colo. If we could find someone at a university that would be willing to host it, that would be ideal. Universities have lots of bandwidth and since this is a mathematical project, they should find it a worthy use of bandwidth. Also, if we were hosting in a college, we wouldn't neccessarily need a 1U case, because we wouldn't get charged by the U (hopefully, we wouldn't get charged at all).

I am a student at PSU. I would be happy to present a proposal to the university. What could we offer a unversity? A university would need to see this as a beneficial project. They are after all financially oriented like a business.

[adpowers]

I don't mind Athlons, I just want to see some mySQL benchmarks on Xeons with dual channel DDR / RDRAM and Athlon MPs.

That Anandtech article I linked to had comparisons of dual Athlons vs. dual Xeons. The Athlons, even while at much slower speeds. I am not sure if they make dual Xeon boards with RDRAM (at least anymore). I think they have switched over to dual DDR because it provides similar bandwidth with lower latency.

That is a valid point. The storage subsystem must fit in the 1U case. How many hard drives can a 1U case contain?

I have seen up to 4 hot swappable hard drives in a 1U case (IDE or SCSI).

How do you plain to do IDE RAID 5? They make controllers but are they integrated into any Dual Athlon MB that fit in a 1/2 U case? Or will the IDE Raid PCI card fit into a xU rack case?

RAID 5 is essential. We can not afford to loose data from a HD crash. Recall that most IDE drives only have a 1 year warranty now. [some high end drives have the 3 year warranty] A 5 year warranty is standard on SCSI drives. A 36GB 10,000 rpm SCSI drive is only $160. SCSI has 15k rmp drives. IDE still does not have any 10k rpm drives.

For IDE RAID 5, you would probably need to build your own, because I haven't seen any (that I can remember of the top of my head) vendors offer IDE RAID 5. Also, the high end IDE hard drives (like the Western Digital SE drives with 8 megs of cache) like I was thinking of using do have 3 year warranties. I think 15k rpm hard drives are unneccesary and too expensive for our needs. Also, I question our need for RAID 5. What is wrong with RAID 1? We would probably be able to get IDE RAID 1 and RAID 1 still provides redundancy.

Since we need service, what are good but cheap Athlon MP rack mount vendors?

If each part comes with a warranty, then why do we need service? You didn't reply to the idea about proxy servers, because if we had them, then service wouldn't be as critical. Also, the second anandtech article I linked to has like 12 vendors that sell the 1U rackmount servers. You can look through them to find a cheap one.

Andrew

[adpowers]

Yes I recall that article. I am a bit lerry of their homegrown DB tests. Woud our DB transactions be similar to a Forum?? What DB system are they using? To confirm their results if they use the DB we want to use (mySQL??) are their any other Athon vs. Xeon benchmarks?

The Athlons do amazingly well at the forum and ad database, and still really well at the website database. Even if the Athlons did the same for our use as they did in the website DB for anandtech, they would still have a better price/performance ratio than the Xeons. You are right about the database, they don't mention what software they use. However, I still think mySQL would preform better on the Athlons.

As for hosting at a university, maybe we could see if curtisc (the highest ranked user on Prime95) would be willing to host it. He is located at Central Missouri State University:
http://www.math-cs.cmsu.edu/~gimps/
Since they already have it running on hundreds of computers and a website dedicated to it, maybe they would be willing to host it. Plus, universities have tons of bandwidth. What is 58 more megs a day? A single kazaa user could surpass that in like an hour.

Andrew

[Prime95]

Originally Posted by adpowers

As for hosting at a university, maybe we could see if curtisc (the highest ranked user on Prime95) would be willing to host it.

An excellent idea. I'll ask him. Saving $100 / month colocation costs would let us invest the savings in a better server.

[ET_]

Originally Posted by Angular

Do you need 1GB of _available_ RAM to cache all of the tablespace?

Well, it obviously depends on the tablespace...

Luigi

[Angular]

For IDE RAID 5, you would probably need to build your own, because I haven't seen any (that I can remember of the top of my head) vendors offer IDE RAID 5. ... Also, I question our need for RAID 5. What is wrong with RAID 1? We would probably be able to get IDE RAID 1 and RAID 1 still provides redundancy.

Software based RAID 1 is not acceptable. (slow, no real time failover)
Hardware based IDE RAID 1 with hot-swap drive carriers is not too bad if that is the price point we need to hit. 2xWD800JB (3y warranty, 8MB 7200 rpm, 80GB) is only ~$220.

RAID 1 gives us the hot backup drive, but will it give us enough speed? We currently have a RAID 5 configuration.

George, are you concerned about HD speed given your DB background? Any concerns about the new server having roughly the same speed storage subsystem speed or lower?

If each part comes with a warranty, then why do we need service? You didn't reply to the idea about proxy servers, because if we had them, then service wouldn't be as critical.

What would our response time be to purchase the needed part and get to the server's location to install it?? This could easily be days not sub 4 hrs. Unless you want to purchase the space parts ahead of time and have a automatic notification of 2-3 competent technicians in the local area that donate their services. If it were a university, we need them to guarantee the server's hardware or pay someone to provide at minimum next day parts and labor.

I am not sure that a next business day service agreement would be enough. A failure that occurs late Friday, is not fixed until Monday?? That really hurts GIMPS's rep and could cause people to quit or not join. Can we afford to be down from Friday to Monday?

The idea of proxy servers is a good idea. However, it is a bad idea to design the reliability of your primary server while counting on a backup proxy that does not exist and may or may not happen. I would like to see a redundant server (or a proxy would be all most as good) in a different location. The security of the proxy's could easily be a problem. Let’s get the primary server configuration decided first without depending on proxies.

Is RH Linux the OS of choice?

Is anyone willing to donate money to this project? Could we setup internet donation system or at least a bank account? How much work is it to make GIMPS a non-profit organization?