NCQ vs TCQ performance?
in [Storage]
|
Prev: notebook SATA drives?
Next: NCQ vs TCQ performance?
From: Tom Forsmo on 6 May 2008 08:42 Arno Wagner wrote: > WD makes this claim for some time now, because they do not have > SCSI disks. I think it is a marketing-lie. If you had read the article I references, you would see they have made measurements comparing the Raptor disk to both SCSI, SATA NCQ and SATA non-NCQ. And the Raptor is by far the fastest disk of the SATA types, for server type load. The results have been feed into a long standing peformance table where you can compare different types of disks, both SCSI and SATA types, if you sort it by performance, you will see SCSI holding the top positions, the highest performant sata disk mentioned on the list is the Raptor. After that you find the all the other SATAs. But my question is, has anything changed with NCQ. But after having studied it, I dont think any changes have occured. The NCQ implementation is really only an implementation of the TCQ simple mode. Simple mode alone gives some performance boost. The referenced article shows that NCQ disk is more suited for desktop load, while TCQ is more suited for server load. And that is why I suspect NCQ has not changed much. regards thomas
From: Arno Wagner on 6 May 2008 11:47 In comp.sys.ibm.pc.hardware.storage Tom Forsmo <spam(a)nospam.net> wrote: > Arno Wagner wrote: >> WD makes this claim for some time now, because they do not have >> SCSI disks. I think it is a marketing-lie. > If you had read the article I references, you would see they have made > measurements comparing the Raptor disk to both SCSI, SATA NCQ and SATA > non-NCQ. And the Raptor is by far the fastest disk of the SATA types, > for server type load. The results have been feed into a long standing > peformance table where you can compare different types of disks, both > SCSI and SATA types, if you sort it by performance, you will see SCSI > holding the top positions, the highest performant sata disk mentioned on > the list is the Raptor. After that you find the all the other SATAs. I did not calim that the Raptor was slow. What WD has claimed, however, time and again, is that the Raptor is just as good as an SCSI disk. It is not. It is, of course, faster than slower spinning (S)ATA drives, unless you are interessted primarily in linear performance. > But my question is, has anything changed with NCQ. But after having > studied it, I dont think any changes have occured. I agree. > The NCQ implementation is really only an implementation of the TCQ > simple mode. Simple mode alone gives some performance boost. The > referenced article shows that NCQ disk is more suited for desktop > load, while TCQ is more suited for server load. And that is why I > suspect NCQ has not changed much. It is also that on high-performance set-ups, the command reordering, buffering, etc. is done by the OS anyways. In extreme cases (e.g. an Oracle server) it is even done by the application. The disk can add a little, but not that much. Arno
From: Tom Forsmo on 7 May 2008 04:35 Arno Wagner wrote: >> What you are saying is that the only reason the raptor is faster is >> because of the spindle speed. which is not correct. > > No, that is not what I am saying. True, latency plays a role, but... > >> Its is also faster because it uses TCQ instead of NCQ. But as >> stated, only for server loads. > > .. as you say. For mostly linear accesses TCQ/NCQ daoe not > play a role and the lower bit density of the Raprors (and > SCSI drives) slow them down to SATA comparable speeds. What do you mean by linear access, do you mean sequential reading/writing (as in reading/writing a large file in one)? > I expect by server loads you mean database loads, webserver > loads and the like, i.e. multiple smaller reads from, > potentially from different processes. Then the queuing > model and latency are certainly or primary interest. By server loads I mostly mean concurrent access by different threads/processes, which includes file access, database and so on. How high the load is, is not as interesting. Both servers and desktops can have a high load, but only servers will have a high concurrency level. So by putting the two comments togheter, what you are saying is that the Raptor is slower because sequntial read/write only as fast or slower than an ordinary NCQ disk? That is correct, but only in it self. The Raptor is designed to be an cheaper entry level server disk, not a desktop disk. So by that, the Raptor is faster, but only for server loads. If you use it in a desktop system its seems to be slower than an ordinary NCQ disk and that is because of the TCQ protocol overhead which in a SATA system does not scale as well as it does in a SCSI system. If one wants really fast disk access no matter what, I would go for a complete SCSI system. But SCSI is 3-5 times more expensive, in addition SCSI do not have as large disks as ATA does, at least which is as cheap. tom
From: Arno Wagner on 7 May 2008 05:50
In comp.sys.ibm.pc.hardware.storage Tom Forsmo <spam(a)nospam.net> wrote: > Arno Wagner wrote: >>> What you are saying is that the only reason the raptor is faster is >>> because of the spindle speed. which is not correct. >> >> No, that is not what I am saying. True, latency plays a role, but... >> >>> Its is also faster because it uses TCQ instead of NCQ. But as >>> stated, only for server loads. >> >> .. as you say. For mostly linear accesses TCQ/NCQ daoe not >> play a role and the lower bit density of the Raprors (and >> SCSI drives) slow them down to SATA comparable speeds. > What do you mean by linear access, do you mean sequential > reading/writing (as in reading/writing a large file in one)? Yes. >> I expect by server loads you mean database loads, webserver >> loads and the like, i.e. multiple smaller reads from, >> potentially from different processes. Then the queuing >> model and latency are certainly or primary interest. > By server loads I mostly mean concurrent access by different > threads/processes, which includes file access, database and so on. How > high the load is, is not as interesting. Both servers and desktops can > have a high load, but only servers will have a high concurrency level. A Linux desktop can have high concurrency at high load without trouble, if the user is compent. That this is rather rare under Windows gets no argument from me. > So by putting the two comments togheter, what you are saying is that the > Raptor is slower because sequntial read/write only as fast or slower > than an ordinary NCQ disk? That is correct, but only in it self. The > Raptor is designed to be an cheaper entry level server disk, not a > desktop disk. Agreed. It is however often mistaken as a blazingly fast desktop "enthusiast" disk, which it is not. > So by that, the Raptor is faster, but only for server loads. Exactly. > If you use it in a desktop system its seems to be slower than an > ordinary NCQ disk and that is because of the TCQ protocol overhead which > in a SATA system does not scale as well as it does in a SCSI system. > If one wants really fast disk access no matter what, I would go for a > complete SCSI system. But SCSI is 3-5 times more expensive, in addition > SCSI do not have as large disks as ATA does, at least which is as cheap. Actually with modern OSes, often getting more RAM wil help a lot. For larger storage needs SCSI/SAS is the way to go. For smaller ones were low throughput on writes is acceptable, FLASH drives may also be an option. If the sustained write rate is low, writes can be speed up by more available RAM. But I agree, if you just want to change one thing, going to SCSI/SAS will give a significant boost with many practical loads. And if you make it RAID with a really good controller, you will also get better throughput (for RAID1 only on reads). Arno |