Drive benchs and pimpage

LiamC

LiamC

0
Joined
Feb 7, 2002
Messages
2,016
Location
Canberra
Hi people, I've been benching some hard drives with some not so common applications and run up a few graphs and thoughts.

This will be sort of a hard drive version of DansData Cooler Guide (if I can make it work).

Anyway, the first part is only a 300BB v D740X v 400JB (a preview because I'm seeking feedback), but I should be adding a 600JB (60GB platter) and 1000BB-SE next week.

http://users.bigpond.net.au/liamc/q42002/hardrives.htm

If you have any comments I'd like to hear them.

I tried to post this in news but kept getting an error
 
who is your audience for this project? I think you could easily make this document work well for people with low-mid level hardware experience. Also, is this intended for people who are interested in buying a new drive how is it going to influence their purchasing decisions?

However, as it stands, you don't seem to be explaining some key things that should be stated for this audience.... Currently, I think you need atleast a good understanding of HDD's to get some of the finer points of your test.


(1) I would explain why you're not listing capacity in the beginning. Explain why the capacity of the individual platters is more important and why it has more of an effect on the performance of a drive...

(2) You might also want to explain the technical things that affect the HDD speed alittle more... you say that "Areal (platter) density is good because we can get more data on and off of the drive faster."

but you don't say what areal densitiy is.... and you don't say that you want more density vs less. and that areal densitiy is related to platter size.


you also say "Larger buffers are good because as with cache on a processor, having the data in a buffer (or cache) is much faster than actually have to seek and read it. "

this is pretty good... but you might want to say that larger buffers are better because they (1) hold recently used data so that if it needs to be retrieved again it can be done so without any mechanical slowdown and (2) provides more oppertunity for read-ahead caching which should also help reduce mechanical slowdowns...

The difference is that in one you're simply saying that a large cache is faster, in the other you're explaining why it is faster and describing it's function.

You don't seem to mention the fact that higher rotational speed leads to faster access times (the only increase in IDE access times we've had in about 5 years was the jump to 7200RPM)





In the end, you add a AMD AthlonXP 1800+ to the mix... You don't seem to state why, you seem to say that you did these tests to see by "how much" the 8MB cahce improves speed... but you dont state this in your conclusion... such as 'The WD's 8mb cache seems to give the same improvement as upgrading to a 300mHz faster processor in business situations.'

stating that "more impact on performance than the jump from 20GB to 40GB platters" is good, but not sufficient for me.

However, I gather that this data may be more useful if you used it to compare two different situations.... someone is buying a PC.. they can choose between the older HDD and a faster processor, or a slower processor and a fast disk.... which to get? well, now they can have you tell em.


on a last note... I don't particualrly care for the graphs.... I would rather eliminate orange because on some monitors the orange square and red square are hard to tell apart on some graphs... Although... I am all in favor of going to a more "prety" graph if you can make em.


well, the design of the site looks good, the data looks good, page loads fast... you've obviously spent some time... I'd like to see your test methods, but they look like they are coming after the experiment... I hope you were consistent.... It also seems fairly well written.. the only thing that stands out to me is the "what we have also seen" repetition near the end.


Overall, I like it, and see that people are probably going to use this or atleast be interested in it.
 
LiamC said:
Hi people, I've been benching some hard drives with some not so common applications and run up a few graphs and thoughts.

This will be sort of a hard drive version of DansData Cooler Guide (if I can make it work).

Anyway, the first part is only a 300BB v D740X v 400JB (a preview because I'm seeking feedback), but I should be adding a 600JB (60GB platter) and 1000BB-SE next week.

http://users.bigpond.net.au/liamc/q42002/hardrives.htm

If you have any comments I'd like to hear them.

I tried to post this in news but kept getting an error
Click to expand...

I got the news you submitted, what error did you get if you don't mind sharing? I was trying to resolve a bug last night with submitting news, but it seemes like a problem continues to exists.
 
blakerwry, I will need a little time to think about your tome :) I'll post again later - had to go shopping with my wife and bub
 
Handruin, the message was this

Not Found
The requested URL /modules/news/ was not found on this server.

Additionally, a 404 Not Found error was encountered while trying to use an ErrorDocument to handle the request.
 
blakerwry,

I do agree with a lot of your feedback - I will expand on some of the points you mentioned, because yes , my site is aimed at the n00b.

The only two points I see differently to you are:

Seek times (IMO) aren't important -> have a look at the SR artile comparing the 1000BB-SE to the 'Cuda IV. Shorter seek times make stuff all difference. Also check out the article on Capacity's effect on server performance -> December 2001.

Also, I did mention why I used an 1800+, just not in the conclusion -> though I will expand on it a little.

As for the graphs, I'll see what I can do, though from UI design, colour is a personal preference so I'm sure I won't please everyone :)

Thanks for the feed back
 
from the 1st article, it shows that in WinBench99 there is only a couple percent difference in speed between the two barracudas, however in IOmeter, there is a ~7% difference in speed....

Access times for the barracuda's differ by .9ms or a ~6% increase in access times when AAM is enabled.


The second article shows how a 10k SCSI drive when only performing half stroke seeks or less can compete with a 15k unit that is performing full stroke seeks or less... since the platter density size is the same, it is apparent that the 15k would normally have an advantage in *BOTH* STR and seek times... by half stroking the 10k drive we virtually eliminate the access time difference and some of the STR difference.


Both of these articles say to me that Access time can make a big difference, going from a 5400 to a 7200 RPM IDE drive makes a difference. Going from a 10k to a 15k makes a difference... Now I would have to agree, that between drives of the same generation a .5ms or so difference in access time is inconsequential in desktop use and debatable in server use.

But since 5400 and 7200 RPM IDE drives are both readily available... both using the newest ATA100 standard, I think you might want to point out that Access times do make a difference if they are significant.
 
Ahh, IOMeter.

SR own doco says that IOMeter readings are more applicable to server loads than personal use PC's. The fact that the 'Cuda made barely any improvement in personal use benchmarks indicates to me that for measuring hard drive performance for personal use, IOMeter is irrelevant (or at least, of greatly reduced significance).

The follow on from that is that (reduced) seek times have little influence on personal use drive performance, a fact that my own testing seems to bear out. Personal use drive performance is largely impacted by transfer rate, cache size and (in IBM's case) the support of Tagged Command Queuing. I come to the IBM conclusion because it supports the same technologies as WD, but also adds TCQ, and according to SR bench's, the IBM's seem to be the fastest IDE drives around.
 
i agree that IOmeter is a better indicator for server use. I also agree that winbench is better indicator for desktop use.

the .9ms access time difference between the 2 barracudas is half the difference between going from a 5400RPM drive to a 7200RPM drive.

If you had 2 drives with the same STR's and same seek times, but one was 5400 and one was 7200 you could expect(from SR's tests) that you would see an aproximate 4% increase in typical desktop perfomance by going to the 7200rpm (not to mention 14% increase in server performance)

wow... that 4% is the same approx. number of difference in your business winstone test between each HDD you tested.... and the same for your Content creation tests in winstone2001 and 2002...


How can you recommend one drive over another if the difference between the drives in your tests is the same as the difference in access times between a 5400RPM and a 7200RPM drive that have otherwise equal specs. (knowing that access times make no noticable difference, in your opinion)


You are basically saying that the difference between these drives is negligable because some one will never be able to notice the speed difference unless they have a benchmark to tell them.
 
now that I think about it, If i sat behind a computer using these drives..

I probably couldn't pick which computers had which drives unless i used a benchmark to tell me.

So, given that, I may as well pick drives on things like reliability, cost, warranty, and of course accoustics.


it's too bad that it's often hard to measure these things...
reliability seems immeasurable at this point...

warranties are somtimes hard to get fullfilled and are often slow...

immediate cost is usually aparent, but a single RMA will cost you the difference between two drives.. not to mention data loss can be more costly than an entire new drive.

accoustics are very subjective... while 2 drives might emit the same level of sound, one might have an annoying pitch. Or idle noise might be the same, but one might have very loud seeks while the other is whisper quiet.
 
blakewry,

One huge difference between my results and SR's is that they use the Winbench reported score - ie the highest. This is psuedo science at its worst. You cannot draw any conclusions from such "golden samples", especially across different benchmarks.

The conclusion you drew above (and SR drew) was that 4% better (lower) access was responsible for the slight gain in Winstone. The golden sample high-score in Winbench probably has no relationship whatsoever to the "golden sample" run in Winstone.

My benchmarks are averages of ten runs in this particular case(though my usual efforts require 25 runs), though I have recently re-run the tests so it may end up being a sample size of 20. With proper statistical analysis of the data, meaningful comparisons can be made. Which I will do when I get time. Not that the access time graph also, is incorrect.

I put together a piece on this here (on Winstone) but it is just as applicable to Winbench.

http://www.realworldtech.com/page.cfm?AID=RWT061902001047
 
LiamC said:
http://www.realworldtech.com/page.cfm?AID=RWT061902001047
Click to expand...
If you remember, I told you about a few mistakes in the graphs of this article and they haven't been corrected yet.

For instance :
winstone-method-fig4.gif


The green line is supposed to be the average of the five best scores. However, we can clearly see that the five highest red quares are located above that green line. How's an average can be lower than all the scores that should compose it?
 
Ah yes Sir Groltz, but you are misinterpreting the results.

The Winstone tests were run in groups of 5. Standard Winstone methodology.

so the first 5 red boxes have a average, and the next 5 and so on.

What you are interpreting is the average of the five highest score (of twenty five). What the green line is, is the average of the five Winstone high scores.

Clearly, from the graph, there is one test of five runs (the one you are pointing out) that is higher than the remaing four test runs (twenty individual runs).

So the graph is consistent and correct, but perhaps my wording could be clearer.
 
Sir Tannin and Lady Tea,

don't look at my drive bench graphs as I ran the D740X with AAM enabled. With it disabled it returned slighty better results, but not much. I also added 1000BB-SE scores, but I haven't had time to upload the graphs. My first attempt lead to horrible pixelated text. But this weekend I am having a break...
 
LiamC said:
Ah yes Sir Groltz, but you are misinterpreting the results.
Click to expand...

Uh...How did I get dragged into this argument?

***Groltz doesn't remember contributing to this thread***
 
Oops :oops:

My sincerest apologies Sir Groltz, I had just read one of your posts and got myself confuzzled. I meant Sir CougTek.
 
LiamC said:
Oops :oops:

My sincerest apologies Sir Groltz, I had just read one of your posts and got myself confuzzled. I meant Sir CougTek.
Click to expand...

No worries, Liam. It could have happened to anyone. :wink:

Might I ask what the royalty titles are for? Being addressed as "Sir" makes my teeth itch.
 
It's more than weird for me to be called "Sir". Myself and the Queen aren't exactly close friends and we don't share much in common from an ideological point of view. You could in fact even say that for the Queen, I'm some kind of Anti-Christ (I'm anti-aristocratic and I despise any type of link between the Crown and Québec : WE ARE INDEPENDANT!).

I would accept to be called Lord CougTek though, as this title isn't only reserved to the British monarchy ;-)

I'm sure this assosciation wasn't meant by LiamC so Steve and I are only increasing our post count here.
 
Lord Groltz und Baron CougTek it shall be. No harm meant, it was just Friday for me, I've been working 60~70 hours a week for a couple of months and I just landed my first weekend off in a while (plus the following Monday). Just going a little stir crazy I think...
 
LiamC said:
Lord Groltz und Baron CougTek it shall be. No harm meant, it was just Friday for me, I've been working 60~70 hours a week for a couple of months and I just landed my first weekend off in a while (plus the following Monday). Just going a little stir crazy I think...
Click to expand...

Nice. Perhaps we can add a von for Cougtek. Baron von Kougtek.
 
In post-count order, the appropriate titles are:

Her Grace the Grand Duchess Tea
His Majesty the Tsar Tannin
His Excellency Baron von Kougtek
The Right Honourable Doug, Duke of Handruin
Il Serenissimo Don Constantine, Professor Emeritus of Wizard
The Right Honorable Viscount Buck
His Lordship Cliptin, Most Noble Marquess of Chatanooga
His Royal Highness Clocker, Crown Prince of Storage
Oh, and I nearly forgot - Comrade Mercutio
 
Tea said:
In post-count order, the appropriate titles are:

Her Grace the Grand Duchess Tea
His Majesty the Tsar Tannin
His Excellency Baron von Kougtek
The Right Honourable Doug, Duke of Handruin
Il Serenissimo Don Constantine, Professor Emeritus of Wizard
The Right Honorable Viscount Buck
His Lordship Cliptin, Most Noble Marquess of Chatanooga
His Royal Highness Clocker, Crown Prince of Storage
Oh, and I nearly forgot - Comrade Mercutio
Click to expand...

So be it, Her Majesty the Grand Duchess of Tea.
 
A Grand Duchess, My Lord (as the daughter of the Tsar) is properly addressed as "Your Grace". In Her Grace Tea's case, however, a simple "hey you" is generally sufficient. Her Grace responds quite positively to "hey you", particularly in when it is included in phrases like "Hey you! Want some chocolate?"
 
Tea said:
In post-count order, the appropriate titles are:

Her Grace the Grand Duchess Tea
His Majesty the Tsar Tannin
His Excellency Baron von Kougtek
The Right Honourable Doug, Duke of Handruin
Il Serenissimo Don Constantine, Professor Emeritus of Wizard
The Right Honorable Viscount Buck
His Lordship Cliptin, Most Noble Marquess of Chatanooga
His Royal Highness Clocker, Crown Prince of Storage
Oh, and I nearly forgot - Comrade Mercutio
Click to expand...

By the way, isn't a Duke the highest order of peerage, and Baron the lowest? I think it is Duke, Marquess, Earl, Viscount, and Baron. Let me check.
 
So, if Höherer Artillerie-Kommandeur von der Sprachwissenschaft Polizei is to be part of this, one should also mention:
His Majesty King Bill the First of Flagreen
LiamC, His Highness Prince Bill the Second
Steve, His Lordship Baron Graf von Groltz
His Holiness Blakewry the Last
and, of course, His Hairiness Jake the Dog
 
How is this: Curly, Most Honorable Order of the Garter. This is the highest order of knighthood. Fittingly, knighthood consists chiefly of orders of chivalry.
 
Jake the Dog said:
yes yes, I know, I'm nothing but hairy pleb. how can I buy my way into the royal court?
Click to expand...

On the contrary, you already have a place in the court. In the tradition of Shakespeare, you provide the services no one else dares suggest. All while wearing suitable disarming identification:


as3461.gif
 
Cliptin said:
Jake the Dog said:
yes yes, I know, I'm nothing but hairy pleb. how can I buy my way into the royal court?
Click to expand...

On the contrary, you already have a place in the court. In the tradition of Shakespeare, you provide the services no one else dares suggest. All while wearing suitable disarming identification:


http://www.clipsahoy.com/clipart2/as3461.gif
Click to expand...
 
You must log in or register to reply here.
Back
Top