Forum Moderators: phranque
I am thinking (well, planning) to replace our current server because it is just no longer up to the task. In the course of my shopping, I have seen some servers with "Dual Processors", or at least that have that upgrade path. I am just curious how these work...
Is this essentially haveing two "computers" in one box, one processor doing one, the other doing the other... or do both processors share the same work? If so, how is it split?
Since we are on the subject of upgrades, any suggestions or hot tips for things to look (or look out) for?
Thanks!
Dave
Most important? Get SCSI
I had two 7200 RPM HDs, and upgraded to 10K RPM 8 MB cache SCSI drives
My read speed went up 1000%!
What I had:
P4 2.0 Ghz
1024 MB RAM
40 GB IDE + 60 GB IDE
now:
Dual Xeon 1.8 Ghz w/ hyperthreading
2048 ECC DDR
3x18.3 GB SCSI + 60 GB IDE
The biggest performance gain I got was from the scsi .. because of the dynamic nature, I was continually accessing the hard disks - as such, reading performance became key [because 2 GB is still not enough]
In your case, if it is very dynamic and will keep growing, I would suggest either getting something like mine, with even a bit more ram, or one simple server with SCSI for MySQL and another box for the actual pages etc.
Also, if you are getting SCSI, get RAID5 ... best performance for database sites .. or even 5/0 if you can afford that! :)
Thats why i bought an Alpha Pal 8045 heatsink (Y.C.Tech 80mm fan) in a Coolermaster Case (six fans), with an Enermax dual fan PSU, it runs quite cold, i can assure you.
carfac,
I've been building computers for years, and they've become so easy to build, that it is negligible what processor, bus tech, and ram type to use. That being said, certainly from the point of cheapness and realiability, i've being using AMD's since 97', and never looked back, certainly back then it was a price issue, as Intel chips were so expensive - however outperforming the AMD's. This cannot be said now, and with the advent of the AMD Athelon XP 'thoroughbred', they are basically wiping the floor of the Intel's.
The biggest problem today faced by computer builders is HEAT. The colder the computer the faster it will run. Personally, getting a high spec MX series dual processing 1.3ghz system, with 3GB of Ram, built on a dual motherboard with built in raid, and large buffer IDE drives, in a very well ventilated case, should be a computer to be proud off, but perform the required task admirrably.
You don't have the bandwidth, or through put on a single server to warrent that kind of storage. Buying multiple drives for a server is a waste of good cash. The perks are in faster cpu's and bigger pipes.
Not necessarily. There are many places that there can be a bottleneck. A competent system administrator can run some tests to find out what's slowing the system down. Low RAM levels and slow disk access are prime suspects on a DB server.
The redundancy factor is mitigated with the newer ultra-reliable drives. If you follow a religious backup procedure nightly or even a couple times a day, I don't see losing too much data. I'd rather have an offsite backup system than spend the money on a raid. I think it all depends on the size of the environment and the frequency of data updating.
Drives will go out. Not as often as the older ones did, but you can bet it will happen. Is a few hours down and a few hours of data lost acceptable? In some applications it is, and the backup strategy you recommend is a good answer. In others it simply is not. A RAID set will let you keep running until you can drop the server off hours and replace the drive -- in under 5 minutes if you're good. And all with no loss of data.
In others it simply is not. A RAID set will let you keep running until you can drop the server off hours and replace the drive -- in under 5 minutes if you're good. And all with no loss of data.
If you are talking about situations where downtime is not acceptable - they've had hot-swap drives for a long time now. There is no need for that five minutes. You just unplug the drive and plug the new one in. The raid starts rebuilding it without stopping regular operations. Performance suffers, but no downtime. Nice cards even blink the lights of the respective drives to make sure you are removing the right one :)
But that's more to the high-end side of things. I believe we are discussing a cost-effective solution here :)
FWIW, I would check out some hardware sites. Post questions in some tech forums - especially when you get down to a component list to buy to put together your server.
Some sites to check out are:
2cpu dot com for dual processor info and motherboard reviews; good forums for asking questions, too
arstechnica dot com -- esp. the cpu & motherboard forum for a critique of your component list prior to purchase
storagereview dot com for reviews on hard drives
anandtech dot com for their stories on their site server upgrades giving real performance figures and benchmarks
pchardware.ro had a couple of articles on building a rackmount server system from scratch a while back
A few items for your consideration:
First, as many have pointed out, getting lots (2-3.5 gigs) of memory would REALLY help a lot. But don't forget you need to set up the server software to cache stuff in memory to cut down on disk access.
As for disk, not sure if it has been pointed out or not, but two advantages of SCSI are:
1) SCSI can handle multiple request at the same time, whereas IDE is a one request at a time, single thread type system. Also, SCSI offloads some of the "housekeeping" from the CPU to the controller card, reducing the load on your CPU for a given amount of disk access/traffic.
2) SCSI are generally regarded as being built to last longer. Just look at the MTBF numbers and check the hardware sites listed above -- esp. storagereview dot com.
To me, THE most important component is the motherboard (with a GREAT quality power supply being VERY important, too.) I would take a look at getting a supermicro or Intel mobo if you go with an Intel CPU.
As for CPUs, I like the price performance of AMD for desktop systems, but if I were putting together a server, I would definitely go Intel at this point. I read the technical papers on AMD and Intel's websites for the AMD and Xeon CPUs. The Xeons seem like a much better CPU. What I really like is that they have ON CHIP circuitry to shut down/slow down the chip if a heat sink fan goes out. This means the response time will be fast enough to save your server from burning up. With the AMD chip, the shut off circuitry is not on the chip -- it's on the motherbord, and only some models at that. Also, I was not comforatable with the response speed of AMD's mobo to CPU communications in the event of needing to shut down the CPU due to rapid temp increase; temps increase VERY rapidly. AMD CPU response time seemed too slow to me, but I don't recall the exact figures -- been a few months.
Nonetheless, lots of folks run AMD servers just fine. Guess it depends on how mission critical your server is to you -- and how good you are at dealing with the more prevelant compatibility issues one deals with with AMD mobo, mobo chipset, PCI card issues (if you run into any) with your OS of choice.
One CPU setup you may consider if heat is an issue to you is a dual P3 1.4 GHZ CPU setup with 512 cache on each CPU. Don't know if that would do the trick for you or not, but it does generate much less heat than either a P4 or an Athlon.
One more point. The reason I would choose an intel CPU on an intel or supermicro SERVER motherboard (besides the on chip Xeon stuff I mentioned above), is due to the intel chipsets on the intel or supermicro SEVER boards. You are much less likely to have flaky problems with an intel cpu, mobo and chipset than you are with an AMD setup. Plus, I don't think the most recent AMD chipsets are setup to work with ECC memory, which one would want with a server IMO. If you do go with AMD, I would look at Tyan and Gigabyte mobo offerings. However, you may be better off to just buy a ready made AMD server since they are likely to have engineered the cooling issues better than what you may end up putting together. I think appro systems are well regarded for AMD server systems.
Note these are MY opinions, others have differing ones they feel stongly about (esp. Intel .vs AMD), so no flames please -- I likely have heard it before and am not going to waste time debating with anyone (wasted enough time doing that before) -- just sharing my opinions for what they are worth; hope they help.
Hope the above helped some. Best wishes on your new server box.
Take care,
Louis
my recommendation would be to maximize the size and type of RAM; and maximize drives (best possible SCSI ultra wide and in good config (RAID 5 for example)), processor would come third place, typically a dual or quad processor isn't nessesary -- that money would be much better spent in the other aspects of the server like more RAM, better SCSI controller, or on redundancies like an RPS. Doing things like duplexing the NIC's and dedicating one for inbound / one for outbound is another way of limiting bottlenecks.
If you are talking about situations where downtime is not acceptable - they've had hot-swap drives for a long time now. There is no need for that five minutes.
True, but that's quite a bit more money to eliminate that 5 minutes. It requires hot-swap drives and hardware RAID, not software. Just stepping up to software RAID on standard drives can cut downtime by hours ... a good step for those who can afford 5 minutes down before they can afford the higher-end hardware.
That plus a removable bay you put a standard HD in and that's it.
The low profile would take smaller space then a beafy single machine. The costs are comparable. When your beafy box fails all dies, on the other hand if you have three boxes even if two fails you can still "limp" along. With load balancing solution, if you need to increase your "powers" just add one more box to the cluster.
I had not checked back in a day or so, and my head is swimming with all the comments!
THANK YOU to you all!
I bought a bit of breathing room... I replaced the celeron 533 in the current box with a Pentium 3 1 GIG, and things are running MUCH better. It was a simple switch of CPU's, pop two dip switches and check the BIOS... and I went from a 66mHz FSB to a 133. Boy and I glad I bought 133 ram! (I wish I could tell you all where it is, because it is much more spunky today!)
Anyway, now I have some time to do some research! This board is GREAT!
dave
last pid: 42552; load averages: 0.15, 0.14, 0.16
This is just coming down off my peak time of day.... yesterday, these were in thre 70-80s!
It's funny how it all works together...
dave
SQL Server, 1 CPU: $4999
SQL Server, 2 CPUs: $9998
Oracle: 1 CPU: $15000
Oracle: 2 CPUs: $30000
So I would max out the processor speed on one CPU before I went to two CPUs.
last pid: 42552; load averages: 0.15, 0.14, 0.16This is just coming down off my peak time of day.... yesterday, these were in thre 70-80s!
This means your CPU is much faster than the rest of the system and is not going to be a bottleneck.
I had to deal with boxes that had 3-5 or even 10-12 average loads during peak times. Now that sucked :)
I had to deal with boxes that had 3-5 or even 10-12 average loads during peak times. Now that sucked :)
Gotcha beat, in the old days before I started using FastCGI to make my code persistant I'd 65+ for at least 6 hours during the day :) But that's just 'cause I was being a dork.
I'll pipe up with my annecdotal recommendations of UMP-v-SMP (and if it matter at all in the grand scheme) and talk about the database server's I have, and the UMP I had, all running MySQL on Linux.
Again, in the olden days (December 2002) I had a DB server that was a single P-III 1gHz with 1GB Ram and some single IDE drive ( I don't remember if it was 5400 or 7200, but i think 7200). I thought it was okay performance-wise until I moved over to my current main DB server, a Dual P-III 750 with 1GB and a 7200 RPM IDE Raid 0+1.
You'll see the obvious difference that the CPUs, while more of them, are slower by a quarter. In fact, in my basic query tests of doing large inserts and large selects, even with the RAID, was slower on the Dual and the Single. However, these tests were just single threads -- when I tried doing a bunch concurrently the Dual machine (even testing with a single disk) did a number on the UMP machine.
My Synopsis: If you're going to using a lot of differnet connections and differing threads, SMP will help, but doing few large queries you won't see much. Of course, you also need an OS that has good SMP capabilities (I use Linux, but I'm told Win2k and FreeBSD have better SMP systems).
I now have two database servers, one primary and one backup. This second machine is a dual P-III 700 with 512MB and a Raid 0 array of two 7200 RPM Seagate Scsi Baraccudas (for speed). However, in the tests I did of these machines ( which was many, many moderately sized queries and inserts ) the 750 blew the 700 away; The 700mHz machine could theorectically hold it's own (slower CPUs but SCSI Drives) but it was getting terrible marks. The keep observer will see the obvious difference: Memory! The other machine had equivilent guts but it was being hobbled by RAM.
Going on this, here are my personal notes about database servers (my own opinions, not gospel):
* UMP for few queries and few clients, SMP for multiple conncurrent queries and clients.
* Memory is most important -- your DB server should never swap
* In order of importance I'd say RAM, Disk and then CPU.
* For disks, all other things being equal, take speed over size if the price is the same because your database isn't THAT big I hope. (e.g: take the 10,000 RPM 18 Gig SCA drive over the 100 Gig IDE 7200)
* The DB should exist on its own drive mechanism so your DB won't have to fight system utilities and other programs for disk seeks
* RAID controllers may look expensive, but your undisturbed sleep is worth more.
* Girls TOTALLY DIG Semetric Multiprocessing (at least, the oned I hang around do).
RAM speed. If RAM is the most important amongst hardware pieces, then between RAM size and speed, speed takes the cake.
Which, of course triggers the next problem, bus speed. Some motherboards have separate bus for memory, some use the same as for the controllers. The latest on this, (for me at least) is the the Intel Northwood core, which does a 4x133Mhz, or about 5.4GB/sec.
So, fastest memory you can get your hands on, matching bus speed, screaming disks, and finaly CPU.
As for screaming disk speeds - for fun I built a RAM drive out of 1MB SIMMs. Even though these SIMMS were only 80ns, that's still much faster then the fastest mechanical hard drive today...(So what if I lost all the data once power is gone? :) )
I think there is an Australian company that makes RAM drives in Gig size, under the name of Wombat or something similar.
then between RAM size and speed, speed takes the cake.
I wouldn't agree. Even the slowest ram is faster than the fastest drive.
For example, you have a table that needs 500Mb of ram to be cached. And you query this table all the time and do sequential scans.
The dbms will just keep it in ram and will return results pretty quickly. But if you have 400Mb of much faster ram, when you need 500Mb - dbms will have to clear it's cache and load the table in two parts - for each query. You'll lose more in disk IO than you gain in quicker memory access.
So if I had to choose, I would pick more cheap ram.
There is no such thing as enough ram for a database. If you can load all your data in ram and still have enough for sort buffers, for all concurrent threads - you'll win big even with slow bus and cheap ram.
