LSI 9260: Write Through or Write Back

[Mitchb]

System: MSI main board, 980X (188 clock, 23 multiplier), 12Gb Dominator 2000 (1880 MHz), 3R0 256gb C300 on LSi 9260 with fastpath using x16 slot.

No read ahead and write through gives CDM on the left. Using read ahead with no other changes gives the middle CDM. Using no read ahead with always write back gives the results on the right.

Using read ahead and write back gives the best throughput, but I loose 4K QD32.

I am open to comments, suggestions, analysis and what not.

Stripe size = 64K.

Thanks

Mitch

[IanB]

Surely it depends what your application is. Are you running a database server? Or is this just a personal machine you're tweaking for the hell of it?

FWIW, if you're not focussed on maxing writes, and most of your needs are down to read speeds (gaming, "normal" computer usage etc.) the middle one seems to have the best overall read performance. Very little "normal" use short of software installation needs writes, that's why the Intel drives (with their massively gimped write speeds) are still real-life contenders in a benchmark-crazy world.

EDIT: did you try 32K stripes, BTW? You may get slightly more performance there. Again, depends on your usage: big file reading (media) then use bigger stripes; small file loading (like system DLLs, apps) you'll get better performance out of smaller stripes, since the data gets spread over more array disks.

[Computurd]

try 16k stripe with c300 on the 9260.
my best results with FP were always
write thru, read ahead, cached i/o.
cached i/o is very important with this.

[Mitchb]

Well, results using 16K stripe with read ahead, cached I/O and write through are below. I have more data that I will post later on.

Mitch

[Computurd]

doing no read ahead usually will boost your 4k32. would be interesting to see the results if you only change that one setting. that 4k single qd is pretty tremendous there though, very impressive. seems to scale well there. that low QD 4k random is really where its at for everyday usage. you might try doing some game loads, and timing them, to see if you will get any difference between the read ahead/no read ahead settings. or maybe some other quantitative performance test with a app you will use in real usage.

[Mitchb]

Two more bench's using 64K stripe.

Mitch

[flamenko]

Why are the 4k randoms so low? Is this a given with the board? I spent last weekend playing with RAID 0 on the Crucials in ICH10 and got my 4ks way up (230 plus) but only after I checked both boxes in the storage policies....specifically turning off windows write cache buffer flushing....