General Conditions

• SCSI are 36 GB 10 rpm (QUANTUM ATLAS_V_36_SCA)
• IDE disks are 40GB 5400 rpm (QUANTUM FIREBALLlct20 40)
• on the same dual Pentum III processor machine, 1 GB of RAM, running linux 7.2, kernel 2.4.9-31smp
• A 8 kByte buffer, XX IO test is done 2 ways : 1 pass single thread pass using only one volume at a time, compared to parralele IO of 10 threads per volume (only one result-average shown in this case (NPV_)).
• XX=5000 and 500 IO shown

• We do not understand (cannot rationalize that is) the dependance on the number of IO and thefore the file size. This is however also seen in the bump observed in the IOZone profile when the file size increases for the same record size. A larger coverage of IOZone is necessary. 40000 IO of 8 kByte falls within the IOZone plateau (rs=3, fs=12). 400000 IO is around fs=15 which is in a valley (before increasing again). We will describe the rest based on 40000 (the largest plateau).
• In write mode, SCSI is 3 to 4 times faster than IDE
• In read mode, the values are comparable but when going to multi stream, IDE performance drops by a factor of 5/6 , SCSI by a factor of 3/4.
• Random seek are, not suprisingly, comparable .
• IOZone profile confirms the difference in write ; the read profile seems the same in unbuffered mode. Since the 2 nodes are equivalent, it seems that the buffered tests are more relevant for the speed difference. There, a factor of 5 in write and read is exhibited for most regions of the phase space record-size/file-size.
• Ignore %tages greater than 100%. They only reflect a system call timing issue (i.e.we relie on the returned value of the time() and clock() functions. Under Linux, the manpages indicates that clock() returns an approximation of processor time used by the program). We will try to address this in future.

• In read mode, since the performance are comparable, IDE being much cheaper, it is reasonable to use it for a distributed disk storage.
• For write mode however, the factor of 3.5 drop moving from SCSI to IDE indicate that a more carefull application 'bound' study needs to be done.

IOPerf results SCSI

		Block IO						Character						Random
		fwrite			fread			putc			getc			seek
Machine/FS	KB	KB/sec Real	KB/sec CPU	%CPU	KB/sec Real	KB/sec CPU	%CPU	KB/sec Real	KB/sec CPU	%CPU	KB/sec Real	KB/sec CPU	%CPU	KB/sec Real	KB/sec CPU	%CPU
_rcas6060r__tmp_jeromel	40000	30769.23	65573.77	46.92	307692.31	285714.29	107.69	9111.62	10025.06	90.89	10389.61	10362.69	100.26	1302.08	1395.09	93.33
rcas6060__tmp_jeromel	400000	13764.62	55788.01	24.67	722.06	137931.03	0.52	1757.16	9498.93	18.50	742.45	9613.07	7.72	1162.57	1162.57	100.00
NPV_rcas6060r__tmp_jeromel	40000	3979.39	38009.29	10.48	79182.75	215062.41	35.53	1834.07	9509.59	19.30	2082.36	10273.07	20.27	287.77	1102.22	28.18
NPV_rcas6060__tmp_jeromel	400000	4261.54	36686.85	11.62	590.85	55328.37	1.07	1907.56	9418.11	20.27	653.13	9444.56	6.92	1015.12	1266.24	79.64

IOPerf results IDE

		Block IO						Character						Random
		fwrite			fread			putc			getc			seek
Machine/FS	KB	KB/sec Real	KB/sec CPU	%CPU	KB/sec Real	KB/sec CPU	%CPU	KB/sec Real	KB/sec CPU	%CPU	KB/sec Real	KB/sec CPU	%CPU	KB/sec Real	KB/sec CPU	%CPU
_rcas6064r__tmp_jeromel	40000	9975.06	65573.77	15.21	285714.29	307692.31	92.86	5571.03	10101.01	55.15	10389.61	10362.69	100.26	1395.09	1395.09	100.00
rcas6064__tmp_jeromel	400000	7171.03	41067.76	17.46	934.43	89086.86	1.05	1015.51	9002.93	11.28	918.32	9378.66	9.79	1061.48	1337.76	79.35
NPV_rcas6064r__tmp_jeromel	40000	1128.66	38742.89	2.92	51301.32	190043.99	26.86	690.81	9450.13	7.31	2090.57	10263.49	20.37	422.67	1080.37	37.60
NPV_rcas6064__tmp_jeromel	400000	1092.82	10190.93	10.83	884.13	70076.66	1.29	1297.84	6122.22	21.33	857.71	9396.88	9.13	905.11	1258.71	72.04

IOzone profiling