INTERVIEW
GIBSON FROM PAGE 61 of any one company like tional RAID than there ure to rebuild every one
sas storage can see and Panasas? has been in at least a of the 1 billion sectors on
share. pNFS is a short My answer then was decade, even though the the failed disk.
form for Parallel NFS, ‘through an industry- low-reliability disk drives Because today’s disk
a new generation of the managed, interoperable oftoday[theoften-scorned array and file system
traditional NFS protocol, and competitive standard Serial ATA desktop disks] technology cannot auto-
properly called NFS Ver- protocol,’ and the pNFS are 10 times more reliable matically cope with the
sion 4.1, that is being spec- proposal was germinated. than the enterprise drives loss of even one sector in
ified in the IETF [Internet Along the way, Los Ala- of a decade ago. 1 billion, these rare disk-
Engineering Task Force] mos and their national And storage vendors read errors become loss
Internet standards body labs friends have been at everywhere are inventing of entire volumes—tera-
and prototyped by compa- the right place and the new marketing names— bytes—ofdata. Sovendors
nies like Panasas, NetApp, right time with a univer- like RAID DP, N+ 3 and are attempting to cope
EMC, IBM, HP, Sun and with all of these prob-
others. lems by employing more
I proposed p NFS to ‘There is less confidence expensive [in capacity and
the NFS Version 4 stan- around solutions based on performance] RAID error-
dards group in late 2003 correcting codes. And
[when PanFS first went traditional RAID than there they are adding on-the-
into production] as a way has been in at least a decade.’ fly testing of checksum
for high-performance codes to notice silent disk
file-system technology errors—much more rare,
to enter into the main- sity research grant, for others—for what RAID but possible.
stream and for main- example, to continue the researchers called RAID 6 Of course the market-
stream NFS technology advancement of pNFS. 15yearsago. Whatisreally ing hype is excessive. I
to be able to provide The bulk of the develop- happening is that the stor- started working on mul-
interoperable solutions ment has and is being age capacity of disks has tiple failure correcting
for high-performance carried by the prototyping been getting bigger at a code for disk arrays in
Linux clusters. companies because the ratethat, onaverage, about 1989, and on disk-array
The pNFS specifica- business case is compel- matches Moore’s Law for technology for reduc-
tion is scheduled to be ling and because leaders the rate of increase of tran- ing rebuild times [called
finished this year, and like Panasas and Los Ala- sistors on computer chips. parity declustering] a few
product announcements mos proposed fair, open That is, today’s disks have years later. Bringing these
are expected not long andevenhandedstandards more than 200 times as technologies into prod-
after that. development processes. much storage capacity. ucts is long overdue, but
But this is about Los What this means is that finally happening across
Alamos and Roadrun- Is there anything else about the time to rebuild a failed the industry.
ner, so you’d be right to the pNFS clustered storage disk is about 200 times When something really
ask for the connection. system that you see as impor- longer, so the window of bad happens—disk read
In fact, the core ideas tant for us to know about vulnerability to second- errors during disk failure
came from a conversa- that hasn’t been discussed ary failures is 200 times rebuilds and maybe a net-
tion with Los Alamos’ here? larger. work error thrown in for
man in the incubation of Yes. Reliability and integ- The probability of the sport—Panasas does not
storage technology, Gary rity. I was one of three disk failing to read back toss away terabytes of data
Grider. Heasked how his authors of the original data is the same as it justbecauseatinyamount
investment in technol- RAID paper in 1988. By was long ago, so today of data is unreachable.
ogy development could 1997, RAID had asserted you can expect at least Instead Panasas automat-
be made persistent—that its dominance in the disk one failed read every ically fences off the file
is, how could he be con- array marketplace and 10TB to 100TB. But the containing problematic
fident that the solutions became the gold standard reconstruction of a failed data and makes the rest
he fosters remain avail- for reliable, high-integrity 500GB disk in an 11-disk of the terabytes of data
able to customers like Los storage. But today there array has to read 5TB, so available to applications
Alamos regardless of the is less confidence around there can be an unaccept- and users without inter-
future product directions solutions based on tradi- ably large chance of fail- ruption. ´
