Self-Healing Single-Bit Errors
Manufacturers of memory modules commonly implement Error-Correcting Code (ECC) algorithms that automatically detect and correct single-bit memory errors (aka bit flips). These errors occur when a single bit of data that should contain a 0 changes to 1, or vice versa. When ECC detects that a bit has an unexpected value, the code can correct the value and proceed.
Microsoft data indicates that as many as 40 percent of -1018 errors are simple bit flips, so implementing an ECC algorithm for the Exchange Store was a no-brainer. Exchange 2003 SP1 does just that and updates the format of pages inside the Exchange database (EDB) from a 4-byte header to an 8-byte header consisting of two 32-bit checksums. The first checksum is an exclusive OR (XOR) checksum that the Store uses to read the correct page off disk or commit the right page into memory. The Store uses the second 32-bit checksum as an ECC check to correct single-bit errors on the page.
Together, the extra intelligence in the Store and the expanded checksum processing lets the Store self-heal some -1018 errors as they occur. To help you track this activity, Exchange uses several event log events. Event ID 474 reports multiple-bit-1018 problems. Event ID 399 indicates that the Store encountered a single-bit error that it was able to correct in memory. (The error might still exist on disk because the Store only corrects a problem on disk when it updates a page and commits that page back into the database.) Event ID 398, which Microsoft believes will be rare, signals a more serious error that occurs when the Store corrects a single-bit error but subsequently believes that the page is still invalid, perhaps because of a checksum failure.
Of course, like any change to the Store, the introduction of ECC slightly complicates the SP1 upgrade procedure. When you apply SP1 to a server, the upgrade process doesn't attempt to upgrade the entire Store at once; doing so would take too long for anything but small servers. Instead, the Store upgrades its page format to accommodate ECC as users access and update individual pages. Over time, the Store will update all the pages to the new checksum format, thus completing the database update. This phased approach means that you can't restore a backup of an SP1 server to a pre-SP1 server because the pre-SP1 Store won't understand the ECC data that SP1 adds to each page and will therefore consider the pages to be corrupt. If you want to update a database completely after installing SP1, you can run Eseutil with the /D parameter to rebuild the database and update each page. Be sure to take a full backup before and after the rebuild, however, because this operation renders invalid any transaction log generated beforehand.
Speaking of backups, Exchange 2003 SP1 also includes a subtle change in the way that the Store processes online backups. The Exchange backup API includes code to validate each page as it streams out of the database into the backup saveset. If you aren't running SP1 and the backup encounters a -1018 error, the backup API terminates processing to avoid a flawed backup. When an online backup is successful, you know that the Exchange database has no detectable problems. Starting in SP1, if a backup finds a single-bit error, it reports the problem but continues with the backup. Therefore, a backup saveset can contain known errors. If you restore the backup, the Store will deal with these errors as it updates the pages. (A multiple-bit error will still cause the backup to fail.)
Worth the Wait
Exchange 2003 SP1 builds on a solid initial release, making some features easier to use, fixing deficiencies in others, and generally tidying up the base product. If you've been waiting for the service pack before you'll upgrade to Exchange 2003, wait no longer. If you're already running Exchange 2003, the decision to upgrade should be straightforward. (The upgrade from Exchange 2003 to SP1 is easy, fast, and requires only one change to the underlying Windows infrastructure—a Microsoft IIS hotfix that affects OWA. The Microsoft article "FIX: IIS 6.0 compression corruption causes access violations," at http://support.microsoft.com/?kbid=831464, explains the problem.) And as an added plus, you can download and use several new tools that offer even more functionality for Exchange 2003 SP1 servers (see the Web-exclusive sidebar "Exchange 2003 SP1Related Tools," http://www.windowsitpro.com, InstantDoc ID 43599, for details). As always, try out the service pack in a test environment and be sure to take a full backup before you upgrade your production servers.
End of Article
