I/O Rates and Write Frequency
As mentioned earlier, write frequency may be expressed in terms
of transactions per day. This is a common practice when describing
processing capabilities of OLTP applications. To apply this write
frequency to our calculations above, we need to convert this into
I/Os per day.
One way of approximating the number of I/Os from transactions per
day of a typical OLTP application is to refer to Transaction Processing
Council's (www.tpc.org) TPC-C benchmark specifications. TPC-C is
a very popular benchmark tool for comparing OLTP performance on
different hardware and software configurations. The unit of performance
measurement of TPC-C is tpmC, which is basically TPC-C transactions
per minute. TPC-C rule of thumb estimate each tpmC (in an optimum
OLTP environment) to correspond to 0.5 I/Os.
1 tpmC = 1 txn/min; 0.5 IOPS per tpmC
1 txn/min = 0.5 IOPS
1 txn = 0.5 IOPS X 1 min X 60 sec/1 min = 30 I/Os
Therefore, TPC-C rules of thumb on tpmC 0.5 IOPS per tpmC = 30 I/Os
per transaction.
Using this derivation, we can translate 2,000,000 transactions a
day to I/Os per day:
I/Os per day = 2,000,000 transactions per day x 30 I/Os per transaction
= 60,000,000 I/Os per day
The computation above covers all the I/Os per day whether they are
read or write I/Os. For our purpose, we are only concerned with
write I/Os and thus have to assume a percentage of this total as
write I/Os.
Most benchmark data on OLTP applications show that 60-70% of total
I/Os are reads. We can therefore estimate write I/Os as:
Write I/Os per day = 40% of 60,000,000 I/Os per day
= 24,000,000 write I/Os per day
RAID Factors
For this example, we will make use of 155 GB E-Disk® Flash drives
in an external RAID subsystem implementing RAID 5 with a minimum
of five disks (2 TB of data will of course consist of at least 13
X 155 GB disks in a single RAID 5 array). There will be a couple
of assumptions we will have to make with this RAID 5 implementation:
First, external subsystem write cache is also disabled. Similar
to that of disabling write cache in an E-Disk® Flash drive,
all write I/Os sent by host to external RAID subsystem writes directly
to disk and does not cache any updates to cache memory of RAID controllers.
RAID will stripe data across all disks in a RAID 5 array. Depending
on RAID 5 implementation, a block of data sent to RAID controllers
may write data anywhere from two disks (data and parity) to all
the disks in an array (worst case). We will assume the worst case
scenario such that any block of data sent to the array results in
writes on all of the disks, thereby wearing out all disks at the
same rate. Note though that this is really a worst-case scenario
and will rarely ever happen, if at all. If, for example, RAID 5
implementation writes to only a third of the total number of disks,
endurance numbers can easily triple simply because of this RAID
factor alone.
I/O Block Size for OLTP Environments
Based on an earlier discussion, I/O block sizes for typical OLTP
environments are small and may range anywhere from 512 Bytes to
8 KB. Since the physical block size in E-Disk® Flash chips is
16 KB, any write I/O with a block size from 512 Bytes to 8 KB will
correspond to only one write in a single physical block in an E-Disk®
drive.
Endurance in Days
To compute for endurance, we will simply have to identify the max.
write I/Os for a 155 GB E-Disk® drive and use write frequency
in terms of I/Os per day.
Max. write cycles for 155 GB E-Disk® drive = 9,961,472,000,000
(*computed earlier in previous section example)
Estimated Write Endurance (days) = 9,961,472,000,000 / 24,000,000
= 415,061 days
Estimated Write Endurance (years) = 415,061 days / (365 days/year)
= 1,137 years Price Trend
Now that we've established the technical viability of solid state
flash disks in mission-critical database applications, let's take
a look at the practicality of flash disk deployment in enterprise
storage. One major thing going for flash solid state disks is the
continued decline in flash memory quotes. The huge market for portable
electronic devices such as mobile phones, digital cameras, and MP3
players has driven semiconductor manufacturers to boost flash memory
density and output. Market research firm Web-Feet Research predicts
that the price per MByte of solid state flash disk will fall by
an average of 80.86 percent annually within a 5-year period starting
2004.
Figure 1: Disk drive comparison, $ / MB
The price gap between HDDs and solid state flash disks is seen to
narrow significantly-with a difference of only about $0.05 per MByte
in 2007-while flash solid state disks dramatically increase in performance
over HDDs (at 100x-150x over in terms of sustained transfers and
IOPS). Web-Feet Research also projects a major decline in the solid
state flash disk/HDD cost ratio by approximately four times, from
433:1 ($0.078 vs. $0.0018 per MByte) in 2003 to 107:1 ($0.096 vs.
$0.0009 per MByte) in 2006. Solid state flash disks will also maintain
its cost-per-MByte advantage over DRAM-based SSDs. Conclusion
Based on the foregoing discussions, we can conclude that flash solid
state disks are good IT investments even in high write environments
such as database applications. The sample case involving BiTMICRO's
E-Disk flash solid state disk (Tables 2 to 6, Sample OLTP environment)
are conservative estimates of actual attainable Erase/Write wear-out
life. We have made several assumptions that drastically increase
the number of writes on E-Disk® Flash drives. For one, we have
disabled the cache. Enabling the E-Disk® SSD cache would definitely
extend endurance in Examples 1 to 2 and Tables 2 to 6 by several
orders of magnitude. Even in our sample OLTP application environment,
we also assumed the worst case RAID 5 scenario where all disks in
a RAID array are written to, notwithstanding the reduction in writes
to the array that RAID cache provides.
Despite these very conservative estimates, the endurance figures
derived are actually very high. Also, if we were to include BiTMICRO's
accelerated Erase/Write endurance verification and testing, E-Disk®
erase/write endurance can be more than 15 times the values in Tables
2 to 6 and Examples 1 and 2. All of these prove that despite the
erase/write endurance limitations, Flash drives are reasonably priced
and can outlast traditional storage devices for practical use in
database and other enterprise applications.
About the Author
Francisco Fronda is currently the Product Manager for BiTMICRO Networks,
Inc. Prior to joining BiTMICRO, he handled sales and technical solutions
for IBM Storage Systems Group. You may contact him at 1 (510) 743-3170
or send an email at francisco.fronda@bitmicro.com.
Page:
1 | 2
| 3
| 4 | BACK
|
