I'm looking at configuring a new server, and I'm having trouble deciding between a SCSI and IDE system. I know that SCSI is considered much better, but I personally have never had a problem using IDE drives in a server. That's why it's difficult to decide.

Anyone feel like persuading me into getting the SCSI system over the IDE one? :)

Go IDE Go. hehehe. SCSI drives hold up against more, so if you're into clients who go all out on looking how stable your setup is, SCSI drives are the way to go. Mind you, IDE drives are plenty stable, SCSI just goes a little further. SCSI makes it a hell of a lot easier to set up a raid config too, if you want, at least it's my opinion that it does. But what do us Canadians know eh?

If you can go with SCSI then go with SCSI without looking at IDE.

Originally posted by Miha
If you can go with SCSI then go with SCSI without looking at IDE.

But why? I've never had a problem with IDE.

If this is for a shared server, SCSI will do you more good since it is much better when it comes to handelling many I/O at the same time than any IDE drive. If the presure on the server is not going to be that high, then IDE will do and you will save a few dollars.

The SCSI drive wasn't too expensive so I decided to go with it. Thanks, all :)

But why? I've never had a problem with IDE.
Faster
Faster
and Faster :)

You made a good choice going with SCSI.

Go with one of those IDE 7200 RPM, 8MB Buffer Hard drives from western digial. I read tons of articles stating they are just as good, if not better than scsi.

www.tomshardware.com (for a review/comparisson)

The issue is not with stability. Modern IDE hard disks are every bit as stable as SCSI hard disks, contrary to what vendors and manufacturers would have you believe (evidence can be found in the fact that IBM rates both its IDE and SCSI hard disks with equal continuous usage reliability). Rather, it is one of performance. IDE hard disks can only process one request at a time, and require the CPU to perform I/O calculations for it. While this is fine for a single use computer, it is hardly acceptable in a server application. SCSI hard disks can process multiple requests at once, and do not require the CPU to perform any calculations on its behalf. Therefore, SCSI makes much more sense for a server application, as it can handle multiple requests simultaneously (which the vast majority of servers are required to do), all while leaving the CPU to calculate other, often critical data (for example, floating-point intensive database calculations are more and more common with the growth of SQL databases). Furthermore, SCSI hard disks are available with much faster access times and sustained (as well as burst) transfer rates, both of which are critical in a typical server application, since servers heavily tax the I/O subsystem. The reason for this is that SCSI hard disks are available with spindle speeds of 10,000 and even 15,000 RPM, whereas the fastest IDE hard disks have only reached 7,200 RPM. This means that the IDE hard disk with the fastest seek time, the Maxtor D740X, with a time of 8ms, pales in comparison to the fastest SCSI drive, the Fujitsu MAM, with a time of 3.5 ms. Furthermore, IDE hard disks typically only have 2mb of cache (excepting Western Digital's Caviar Special Edition hard disks, with 8mb of cache), and therefore cannot store as much frequently accessed data. The majority of SCSI hard disks have 8mb of cache, and a significantly higher storage to cache ratio. The advantage with IDE disks is simple: cost. An IBM Deskstar 120GXP 120gb, arguably the best IDE hard disk, can be had for just over $100. An 18.3gb Fujitsu MAM, on the other hand, costs closer to $220. The best 10,000rpm SCSI disk, the Maxtor Atlas 10k III, costs over $150 for an 18.3gb version. Another issue, however, is that none of these high performance 10k and 15k rpm SCSI disks are available in capacities greater than 73.4gb. Finally, to SCSI's benefit, SCSI's performance under RAID is phenomenal. Since IDE hard disks can only process one request at a time and require CPU assistance, they don't respond as well in RAID arrays and increase CPU utilization phenomenally (for example, my home computer, which runs two Western Digital WD1200SJ hard disks in RAID 0, sometimes sees CPU utilizations upwards of 40%). SCSI, on the other hand, merely multiplies its positive qualities. Essentially, it's an issue of cost. If you can afford SCSI, get it. If not, go with IDE.

Originally posted by Daovonnaex
The issue is not with stability. Modern IDE hard disks are every bit as stable as SCSI hard disks, contrary to what vendors and manufacturers would have you believe (evidence can be found in the fact that IBM rates both its IDE and SCSI hard disks with equal continuous usage reliability). [/i]


Some IDE drive manufacturers use a higher viscosity lube on the spindle for the desktop IDE drive so that they spin up more quickly. You may want to avoid the more "desktop-oriented" IDE drives for this reason.

Rather, it is one of performance. IDE hard disks can only process one request at a time, and require the CPU to perform I/O calculations for it. While this is fine for a single use computer, it is hardly acceptable in a server application. SCSI hard disks can process multiple requests at once, and do not require the CPU to perform any calculations on its behalf.


Careful here. Newer IDE drives (such as some from IBM) implement tagged i/o requests. FreeBSD has added tagged i/o to the ATA driver. Additionally, some IDE controller cards, such as those from 3ware (http://www.3ware.com) have on-board processors that remove the processing burden from the main CPU(s). The 3Ware cards are rather high-performing. The 3Ware IDE card uses the SCSI miniport driver, so the CPU utilization is the same as that of a SCSI drive.

Therefore, SCSI makes much more sense for a server application, as it can handle multiple requests simultaneously (which the vast majority of servers are required to do),

That depends upon both the load and the usage of the server.


... leaving the CPU to calculate other, often critical data (for example, floating-point intensive database calculations are more and more common with the growth of SQL databases).

The Intel CPUs have a separate floating point processor, separate from the main CPU processor that would be used for IDE processing.



Furthermore, SCSI hard disks are available with much faster access times and sustained (as well as burst) transfer rates, both of which are critical in a typical server application, since servers heavily tax the I/O subsystem.


Yup, access times are faster. Sustained data throughput rates are not necessarily faster.



The reason for this is that SCSI hard disks are available with spindle speeds of 10,000 and even 15,000 RPM, whereas the fastest IDE hard disks have only reached 7,200 RPM. This means that the IDE hard disk with the fastest seek time, the Maxtor D740X, with a time of 8ms, pales in comparison to the fastest SCSI drive, the Fujitsu MAM, with a time of 3.5 ms.

Plus, there seems to be a bit more overhead in IDE command processing, adding more to the overall access times. This becomes important if your server is doing a lot of database-type activities. For most web-server stuff, it may not even be noticeable.


Furthermore, IDE hard disks typically only have 2mb of cache (excepting Western Digital's Caviar Special Edition hard disks, with 8mb of cache), and therefore cannot store as much frequently accessed data. The majority of SCSI hard disks have 8mb of cache, and a significantly higher storage to cache ratio.


So buy the IDE drive with the 8MB cache. duh....


The advantage with IDE disks is simple: cost. An IBM Deskstar 120GXP 120gb, arguably the best IDE hard disk, can be had for just over $100. An 18.3gb Fujitsu MAM, on the other hand, costs closer to $220. The best 10,000rpm SCSI disk, the Maxtor Atlas 10k III, costs over $150 for an 18.3gb version. Another issue, however, is that none of these high performance 10k and 15k rpm SCSI disks are available in capacities greater than 73.4gb.


You should be fairer in your cost comparisons. Throughout most of your entire prior argument, you were using the performance capabilities of the 15krpm SCSI drives to shoot down the IDE drives. So you should use the cost of the 15krpm drives as well, not the cheaper 10krpm drives.


Finally, to SCSI's benefit, SCSI's performance under RAID is phenomenal. Since IDE hard disks can only process one request at a time and require CPU assistance, they don't respond as well in RAID arrays and increase CPU utilization phenomenally (for example, my home computer, which runs two Western Digital WD1200SJ hard disks in RAID 0, sometimes sees CPU utilizations upwards of 40%).

Is that software RAID? If so, no wonder it is so CPU intensive. Take a look at the 3Ware site, specifically the 7000 series (http://www.3ware.com/products/escalade.asp). That is IDE RAID done right. And when you throw in the lower cost of the IDE drives, the 3Ware IDE solution is actually less expensive than the lower-capacity, lower-performing SCSI solution.


SCSI, on the other hand, merely multiplies its positive qualities.

An unsubstantiated ambiguous assertion.


Essentially, it's an issue of cost. If you can afford SCSI, get it. If not, go with IDE.

Not really. If IDE meets your needs, why pay more for SCSI? I used to be a SCSI-bigot until the IDE hardware RAID controllers came along. Now I reserve SCSI only for intensive DBMS applications on my servers.

Save the money you would spend on SCSI and buy more system memory, there may be more of a performance improvement there.

Originally posted by Mike the newbie

Careful here. Newer IDE drives (such as some from IBM) implement tagged i/o requests. FreeBSD has added tagged i/o to the ATA driver. Additionally, some IDE controller cards, such as those from 3ware (http://www.3ware.com) have on-board processors that remove the processing burden from the main CPU(s). The 3Ware cards are rather high-performing. The 3Ware IDE card uses the SCSI miniport driver, so the CPU utilization is the same as that of a SCSI drive.
[/B]I wasn't aware that IBM's latest ATA drives implemented tagged I/O requests, though it is only logical, since IBM is the only company that uses ATA hard disks in large-scale media arrays. I am pleased to see that for the first time since IBM's MCA architecture that someone has implemented a processor that remove the burden from the CPU.


Originally posted by Mike the newbie
The Intel CPUs have a separate floating point processor, separate from the main CPU processor that would be used for IDE processing.You're dreaming. The last CPU Intel sold with a math coprocessor was the 80286. All Intel processors since the 386 have had an integrated fpu. Furthermore, since the P6 architecture debuted (Pentium Pro, Pentium II, Pentium III, and all Celerons but the latest Willamette-128 core Celerons), Intel processors have converted instructions from x86 code into ROPs (RISC Operation Processes) and then back into x86 code. The P7 (Pentium 4, Xeon, and latest Celeron) and the K7 (Athlon, Duron) do the same, albeit more efficiently. A CPU can only convert a limited number of instructions at once, and therein lies the danger with have to calculate IDE functions in a database server.



Originally posted by Mike the newbie
Yup, access times are faster. Sustained data throughput rates are not necessarily faster.The IDE hard disks with the fastest sustained data rates, the IBM Deskstar 60GXP and 120GXP, have a sustained transfer rate of just over 40 MB/second. The fastest SCSI disks for transfer (Maxtor Atlas 10K III, Fujitsu MAN, Seagate Cheetah 73LP, Fujitsu MAM, and Seagate Cheetah X15-36LP) average just over 60 MB/second.

Originally posted by Mike the newbie
So buy the IDE drive with the 8MB cache. duh....The slower DRAM used in the Special Edition Wester Digital drives and the higher storage/cache ratio still mean that a SCSI disk with 8MB of cache is superior.


Originally posted by Mike the newbie
You should be fairer in your cost comparisons. Throughout most of your entire prior argument, you were using the performance capabilities of the 15krpm SCSI drives to shoot down the IDE drives. So you should use the cost of the 15krpm drives as well, not the cheaper 10krpm drives.I mentioned the cost of a Fujitsu MAM, which is a 15k rpm SCSI drive. I further cite to cost as a downside.


Originally posted by Mike the newbie
Is that software RAID? If so, no wonder it is so CPU intensive. Take a look at the 3Ware site, specifically the 7000 series (http://www.3ware.com/products/escalade.asp). That is IDE RAID done right. And when you throw in the lower cost of the IDE drives, the 3Ware IDE solution is actually less expensive than the lower-capacity, lower-performing SCSI solution.It's on onboard Promise Controller of my ABIT TH7II-RAID (it offloads all calculations to the CPU, so it is technically software RAID). Since I have a heavily, heavily overclocked Pentium 4 (and this is not a server, mind you), the CPU utilization does not bother me (especially since software RAID 0 technically outperforms hardware RAID 0).


Originally posted by Mike the newbie
An unsubstantiated ambiguous assertion.Not necessarily. There's a reason that no company other than IBM uses ATA hard disks for large scale media arrays, and it's due to the lack of tagged I/O requests on IDE disks.




Originally posted by Mike the newbie
Not really. If IDE meets your needs, why pay more for SCSI? I used to be a SCSI-bigot until the IDE hardware RAID controllers came along. Now I reserve SCSI only for intensive DBMS applications on my servers.

Save the money you would spend on SCSI and buy more system memory, there may be more of a performance improvement there. A very valid point indeed. I did include the disclaimer regarding affordability, but added memory would indeed aid server performance more that SCSI hard disks in the vast majority of server applications.

According to this article (http://www.tecchannel.de/hardware/964/index.html (http://www.tecchannel.de/hardware/964/index.html)) on TechChannel the IDE disk aren't built for continious use.

I think that more than a reasonable argument to go for SCSI.

It seems that most of the IDE disk are calculated for a 5 hour a day ratio. Not the 750 hourse that a normal month has :)

Originally posted by dot.K
According to this http://www.tecchannel.de/hardware/964/index.html

The writer of the article obiously didn't know what MTBF means.
And 5 hours per day? So everyone on a office pc only works from 9:00 to 14:00 ?
:D

I'd argue with you on the 120GXP as even IBM don't recommend it be turned on for more than 8 hours at a time. As for the sustained transfer, you'll find that the fastest IDE drives can do about 56mb/s from the outer tracks and SCSI drives about 70-80mb/s from the outer tracks due to the higher spindle speeds.

We use the 3Ware and Adaptec IDE RAID cards and they give a much better price/performance ration than SCSI.

We're big fans of the 3Ware cards. We recently just put up a server with 3 80GB drives (2 in a mirrored array, 1 as a hot standby), and saved at least $2500 on the cost of the machine when compared with using 73GB SCSI drives, and got an additional 7GB of storage :)

For the SCSI pundits .... 99% of the web servers out there don't even push 1Mb/sec. Considering that even the crappiest of IDE drives can manage to pump out a measly 20MB/sec, you won't have any problems throughput wise using IDE drives. Hell, you could probably use a ZIP drive and not have any real disk access problems.

Regardless, redundancy is what's really important to me. Ultimately, it's my job and responsiblity to keep the server up and running, no matter what. I've been down the SCSI RAID5 route, and I wasn't impressed at all. Maybe I just had a string of bad luck, but in a matter of 2 months I had 4 or 5 SCSI drives die causing me to completely loose a RAID5 array twice. I wasn't happy, my customers weren't happy and a ton of man hours were wasted rebuilding the machine / restoring from backups.

Cost wise and redundency wise it makes all the sense in the world to setup a mirrored array with a hot stand by drive, using IDE. With this setup, I effectively have N+2 redundancy. I know I can sleep soundly at night. I also know that I managed to save $2500 minimum - money that I can spend on other equipment or software, which will continue to improve the services we offer. I also save 1U of rack space, which adds up to a couple hundred bucks of year in savings.

Another good point is to really load the server up on RAM. If you have to deal with a huge swap file / disk trashing, I don't care what type of drives you have - your server is going to slow to a crawl.

And that's just my 2 cents :)

Originally posted by Daovonnaex
You're dreaming. The last CPU Intel sold with a math coprocessor was the 80286. All Intel processors since the 386 have had an integrated fpu. Furthermore, since the P6 architecture debuted (Pentium Pro, Pentium II, Pentium III, and all Celerons but the latest Willamette-128 core Celerons), Intel processors have converted instructions from x86 code into ROPs (RISC Operation Processes) and then back into x86 code. The P7 (Pentium 4, Xeon, and latest Celeron) and the K7 (Athlon, Duron) do the same, albeit more efficiently. A CPU can only convert a limited number of instructions at once, and therein lies the danger with have to calculate IDE functions in a database server.

I wasn't dreaming. I was speaking of the logical function, not the physical 8087 or 80287 chip. There was, and still is, a separate floating point processor in all the Intel CPUs.

But this is a non-issue because the 3Ware IDE controller appears to the OS as a SCSI drive, not an IDE drive.



The IDE hard disks with the fastest sustained data rates, the IBM Deskstar 60GXP and 120GXP, have a sustained transfer rate of just over 40 MB/second. The fastest SCSI disks for transfer (Maxtor Atlas 10K III, Fujitsu MAN, Seagate Cheetah 73LP, Fujitsu MAM, and Seagate Cheetah X15-36LP) average just over 60 MB/second.

The slower DRAM used in the Special Edition Western Digital drives and the higher storage/cache ratio still mean that a SCSI disk with 8MB of cache is superior.

It is not necessarily a matter of superiority, but a matter of applicability. Hosting is a business, and to run a business you need to be very much aware of price/benefit ratios and things like return on investment. To spend money where it is not needed is not good business sense.

Whenever I see the recommendation "if you can afford SCSI, buy it", my opinion of the person giving that advice diminishes because a recommendation is being made without knowing the business needs. There may be some applications where plain IDE drives are more than appropriate.

Indeed, I suspect that the 3Ware IDE RAID controllers could outperform SCSI RAID in some applications (that is, until Ultra320 SCSI becomes more popular). I've seen sustained data rates of over 170MBytes per second coming off the 3Ware RAID controllers. That's a lot of concurrent media streams. And when you consider the much lower cost per megabyte of IDE drives (and the significantly lower cooling per megabyte required for IDE drives), you have to start wondering when SCSI RAID is actually needed.

At this point, I'm recommending SCSI RAID only for the most demanding server applications. The money saved is better used buying memory and even a faster CPU (or adding a second CPU). I suspect the overwhelming majority of SCSI drives "out there" are very much under-utilized.

On the other hand, if you have money to burn (in the case of SCSI heat dissipation needs, that may be taken literally:)), then buy SCSI drives without even thinking if you really need them.

Originally posted by KDAWebServices
I'd argue with you on the 120GXP as even IBM don't recommend it be turned on for more than 8 hours at a time....


Can you point me to a reference on IBM's web site to substantiate that assertion?

There was some misinformation posted in a Storagereview article at one point, but that info was later retracted.

More info is here. (http://www.theinquirer.net/?article=2926)