MMC double buffering
Per Forlin
per.forlin at linaro.org
Fri Dec 17 15:38:23 UTC 2010
Hi again,
I made a mistake in my double buffering implementation.
I assumed dma_unmap did not do any cache operations. Well, it does.
Due to L2 read prefetch the L2 needs to be invalidated at dma_unmap.
I made a quick test to see how much throughput would improved if
dma_unmap could be run in parallel.
In this run dma_unmap is removed.
Then the figures for read becomes:
* 7-16 % gain if double buffering in the ideal case. Closing on the
same performance as for PIO.
Relative diff: MMC-VANILLA-DMA-LOG -> MMC-MMCI-2-BUF-DMA-LOG-NO-UNMAP
CPU is abs diff
random random
KB reclen write rewrite read reread read write
51200 4 +0% +0% +7% +8% +2% +0%
cpu: +0.0 +0.0 +0.7 +0.7 -0.0 +0.0
51200 8 +0% +0% +10% +10% +6% +0%
cpu: -0.1 +0.1 +0.6 +0.9 +0.3 +0.0
51200 16 +0% +0% +11% +11% +8% +0%
cpu: -0.0 -0.1 +0.9 +1.0 +0.3 +0.0
51200 32 +0% +0% +13% +13% +10% +0%
cpu: -0.1 +0.0 +1.0 +0.5 +0.8 +0.0
51200 64 +0% +0% +13% +13% +12% +1%
cpu: +0.0 +0.0 +0.4 +1.0 +0.9 +0.1
51200 128 +0% +5% +14% +14% +14% +1%
cpu: +0.0 +0.2 +1.0 +0.9 +1.0 +0.0
51200 256 +0% +2% +13% +13% +13% +1%
cpu: +0.0 +0.1 +0.9 +0.3 +1.6 -0.1
51200 512 +0% +1% +14% +14% +14% +8%
cpu: -0.0 +0.3 +2.5 +1.8 +2.4 +0.3
51200 1024 +0% +2% +14% +15% +15% +0%
cpu: +0.0 +0.3 +1.3 +1.4 +1.3 +0.1
51200 2048 +2% +2% +15% +15% +15% +4%
cpu: +0.3 +0.1 +1.6 +2.1 +0.9 +0.3
51200 4096 +5% +3% +15% +16% +16% +5%
cpu: +0.0 +0.4 +1.1 +1.7 +1.7 +0.5
51200 8192 +5% +3% +16% +16% +16% +2%
cpu: +0.0 +0.4 +2.0 +1.3 +1.8 +0.1
51200 16384 +1% +1% +16% +16% +16% +4%
cpu: +0.1 -0.2 +2.3 +1.7 +2.6 +0.2
I will work on adding unmap to double buffering next week.
/Per
On 16 December 2010 15:15, Per Forlin <per.forlin at linaro.org> wrote:
> Hi,
>
> I am working on the blueprint
> https://blueprints.launchpad.net/linux-linaro/+spec/other-storage-performance-emmc.
> Currently I am investigating performance for DMA vs PIO on eMMC.
>
> Pros and cons for DMA on MMC
> + Offloads CPU
> + Fewer interrupts, one single interrupt for each transfer compared to
> 100s or even 1000s
> + Power save, DMA consumes less power than CPU
> - Less bandwidth / throughput compared to PIO-CPU
>
> The reason for introducing double buffering in the MMC framework is to
> address the throughput issue for DMA on MMC.
> The assumption is that the CPU and DMA have higher throughput than the
> MMC / SD-card.
> My hypothesis is that the difference in performance between PIO-mode
> and DMA-mode for MMC is due to latency for preparing a DMA-job.
> If the next DMA-job could be prepared while the current job is ongoing
> this latency would be reduced. The biggest part of preparing a DMA-job
> is maintenance of caches.
> In my case I run on U5500 (mach-ux500) which has both L1 and L2
> caches. The host mmc driver in use is the mmci driver (PL180).
>
> I have done a hack in both the MMC-framework and mmci in order to make
> a prove of concept. I have run IOZone to get measurements to prove my
> case worthy.
> The next step, if the results are promising will be to clean up my
> work and send out patches for review.
>
> The DMAC in ux500 support to modes LOG and PHY.
> LOG - Many logical channels are multiplex on top of one physical channel
> PHY - Only one channel per physical channel
>
> DMA mode LOG and PHY have different latency both HW and SW wise. One
> could almost treat them as "two different DMACs. To get a wider test
> scope I have tested using both modes.
>
> Summary of the results.
> * It is optional for the mmc host driver to utitlize the 2-buf
> support. 2-buf in framework requires no change in the host drivers.
> * IOZone shows no performance hit on existing drivers* if adding 2-buf
> to the framework but not in the host driver.
> (* So far I have only test one driver)
> * The performance gain for DMA using 2-buf is probably proportional to
> the cache maintenance time.
> The faster the card is the more significant the cache maintenance
> part becomes and vice versa.
> * For U5500 with 2-buf performance for DMA is:
> Throughput: DMA vanilla vs DMA 2-buf
> * read +5-10 %
> * write +0-3 %
> CPU load: CPU vs DMA 2-buf
> * read large data: minus 10-20 units of %
> * read small data: same as PIO
> * write: same load as PIO ( why? )
>
> Here follows two of the measurements from IOZones comparing MMC with
> double buffering and without. The rest you can find in the text files
> attached.
>
> === Performance CPU compared with DMA vanilla kernel ===
> Absolute diff: MMC-VANILLA-CPU -> MMC-VANILLA-DMA-LOG
> random random
> KB reclen write rewrite read reread read write
> 51200 4 -14 -8 -1005 -988 -679 -1
> cpu: -0.0 -0.1 -0.8 -0.9 -0.7 +0.0
>
> 51200 8 -35 -34 -1763 -1791 -1327 +0
> cpu: +0.0 -0.1 -0.9 -1.2 -0.7 +0.0
>
> 51200 16 +6 -38 -2712 -2728 -2225 +0
> cpu: -0.1 -0.0 -1.6 -1.2 -0.7 -0.0
>
> 51200 32 -10 -79 -3640 -3710 -3298 -1
> cpu: -0.1 -0.2 -1.2 -1.2 -0.7 -0.0
>
> 51200 64 +31 -16 -4401 -4533 -4212 -1
> cpu: -0.2 -0.2 -0.6 -1.2 -1.2 -0.0
>
> 51200 128 +58 -58 -4749 -4776 -4532 -4
> cpu: -0.2 -0.0 -1.2 -1.1 -1.2 +0.1
>
> 51200 256 +192 +283 -5343 -5347 -5184 +13
> cpu: +0.0 +0.1 -1.2 -0.6 -1.2 +0.0
>
> 51200 512 +232 +470 -4663 -4690 -4588 +171
> cpu: +0.1 +0.1 -4.5 -3.9 -3.8 -0.1
>
> 51200 1024 +250 +68 -3151 -3318 -3303 +122
> cpu: -0.1 -0.5 -14.0 -13.5 -14.0 -0.1
>
> 51200 2048 +224 +401 -2708 -2601 -2612 +161
> cpu: -1.7 -1.3 -18.4 -19.5 -17.8 -0.5
>
> 51200 4096 +194 +417 -2380 -2361 -2520 +242
> cpu: -1.3 -1.6 -19.4 -19.9 -19.4 -0.6
>
> 51200 8192 +228 +315 -2279 -2327 -2291 +270
> cpu: -1.0 -0.9 -20.8 -20.3 -21.0 -0.6
>
> 51200 16384 +254 +289 -2260 -2232 -2269 +308
> cpu: -0.8 -0.8 -20.5 -19.9 -21.5 -0.4
>
> === Performance CPU compared with DMA with MMC double buffering ===
> Absolute diff: MMC-VANILLA-CPU -> MMC-MMCI-2-BUF-DMA-LOG
> random random
> KB reclen write rewrite read reread read write
> 51200 4 -7 -11 -533 -513 -365 +0
> cpu: -0.0 -0.1 -0.5 -0.7 -0.4 +0.0
>
> 51200 8 -19 -28 -916 -932 -671 +0
> cpu: -0.0 -0.0 -0.3 -0.6 -0.2 +0.0
>
> 51200 16 +14 -13 -1467 -1479 -1203 +1
> cpu: +0.0 -0.1 -0.7 -0.7 -0.2 -0.0
>
> 51200 32 +61 +24 -2008 -2088 -1853 +4
> cpu: -0.3 -0.2 -0.7 -0.7 -0.2 -0.0
>
> 51200 64 +130 +84 -2571 -2692 -2483 +5
> cpu: +0.0 -0.4 -0.1 -0.7 -0.7 +0.0
>
> 51200 128 +275 +279 -2760 -2747 -2607 +19
> cpu: -0.1 +0.1 -0.7 -0.6 -0.7 +0.1
>
> 51200 256 +558 +503 -3455 -3429 -3216 +55
> cpu: -0.1 +0.1 -0.8 -0.1 -0.8 +0.0
>
> 51200 512 +608 +820 -2476 -2497 -2504 +154
> cpu: +0.2 +0.5 -3.3 -2.1 -2.7 +0.0
>
> 51200 1024 +652 +493 -818 -977 -1023 +291
> cpu: +0.0 -0.1 -13.2 -12.8 -13.3 +0.1
>
> 51200 2048 +654 +809 -241 -218 -242 +501
> cpu: -1.5 -1.2 -16.9 -18.2 -17.0 -0.2
>
> 51200 4096 +482 +908 -80 +82 -154 +633
> cpu: -1.4 -1.2 -19.1 -18.4 -18.6 -0.2
>
> 51200 8192 +643 +810 +199 +186 +182 +675
> cpu: -0.8 -0.7 -19.8 -19.2 -19.5 -0.7
>
> 51200 16384 +684 +724 +275 +323 +269 +724
> cpu: -0.6 -0.7 -19.2 -18.6 -19.8 -0.2
>
More information about the linaro-dev
mailing list