Performance results of TSS device

Performance results of TSS with different configurations is summarized in Table 3. Trace runs of TSS was also done with different policies and the results summarized in Table 4. Note that we are not comparing our results with the existing RAID in Linux, as our main aim here is to consider the advantages/disadvantages of TSS with/without policies and not comparing the different static RAID systems.

Table 3: Performance of RAID devices through TSS

Disk No.	Percentage	Type of	Read Time	Write Time	I/O Time	Space
	of Writes	Device	per Blk (ms)	per Blk (ms)	per Blk (ms)	overhead
5 (/usr)	59.83	RAID0	4.40	7.17	6.70	100%
		RAID1	5.89	8.95	8.44	250%
		RAID5	4.94	20.91	18.22	125%
		cRAID5	29.74	106.81	93.84	87.2%
6 (/var)	83.17	RAID0	4.30	6.98	5.90	100%
		RAID1	4.94	8.45	7.04	250%
		RAID5	4.76	19.32	13.47	125%
		cRAID5	23.70	74.36	54.00	87.5%

Table 4: Performance of TSS device with policies

Disk No.	Percentage	Type of	Read Time	Write Time	I/O Time	Space
	of Writes	Device	per Blk (ms)	per Blk (ms)	per Blk (ms)	overhead 2
5 (/usr)	59.83	TSS with	6.41	28.03	24.39	205%
		watermarks
		TSS with	6.06	13.35	12.12	237%
		5 min rule
6 (/var)	83.17	TSS with	7.77	26.32	18.87	190%
		watermarks
		TSS with	5.43	14.53	10.87	225%
		5 min rule

From Table 3 we find that the access times in RAID0 is the least, and cRAID5 is the highest. This result is on the expected lines. Notable result here is RAID5 to RAID1 ratio in terms of access time is around 2 and in terms of space is 0.5. cRAID5 access time is around four to five times that of RAID5, and uses around 60% of the space used by RAID5.

RAID0 is just 1.2 times faster than RAID1, mainly because RAID0 does not have the benefit of asynchronous writes in it. RAID1 performance matches with that of RAID0, even with additional redundancy and logging features in it.

RAID5 access times is nearly twice that of RAID1. This can be attributed to:

• RMW cycles for short writes
• RAID5 has an additional parity calculation time compared to RAID1.
• Due to asynchronous writes, there is no time spent waiting for the disks to complete the writes. So, the additional write request in case of RAID1 does not reduce performance.

cRAID5 access times is much higher than RAID5. This can be attributed to:

• Compression / Decompression of the whole stripe needed even for small data requests.
• cRAID5 involves lot of copying from buffer cache to page cache, during compression / decompression. This is due to the fact that the page cache and buffer cache are separate in Linux-2.2.5.

These comparative figures gives good indications for setting the policy parameters in an integrated TSS device. On the basis of performance, the RAID1 and RAID5 parts should have around the same number of stripes and cRAID5 a smaller number of stripes. This also can vary depending on other things like available free space in the device.

The trace runs with watermark policy was carried out with the following watermarks:
RAID1 - Lower Watermark 20-30%, Upper 50-60%
RAID5 - Lower Watermark 25-35%, Upper 55-65%
cRAID5 - Lower Watermark 5-10%, Upper 20-25%

The trace run with the five minute rule[8] was done with RAID1-RAID5 transition frequency as 20 seconds and RAID5-cRAID5 transition frequency as 350 seconds. The actual values got from five-minute rule formula have not been used in case of trace runs, as the execution does not happen in real time but compacted in time. So the actual RAID1-RAID5 and RAID5-cRAID5 transition frequencies from five-minute rule had to be reduced by a large factor.

TSS with watermarks has access time more than that of normal RAID5. This is mainly due to the static nature of watermark policy. If all watermarks are conforming, and no new blocks are accessed, then it does not do any more migrations, and the temperatures of the stripes are not used at all. This can result in a severe performance penalty.

TSS with five-minute rule has access time in between that of RAID1 and RAID5. This improvement in performance compared to watermarks is mainly due to dynamic stripe placements that happen on a continuous manner depending on the temperature.